Projects

Case studies from real deployments.

49 projects across 13 industries and 13 business functions. Filter by the sector you are in, or by the part of the business you are trying to fix - a workflow proven in one industry is usually the same workflow in yours.

Industries

Functions

49 projects

AI-Assisted Enquiry Management

Top-Tier Macroeconomic Policymaker

A drafting assistant grounded in the organization's own precedents saved at least €56k of manual effort while holding to corporate tone.

#Government
#Administration - Correspondence

Problem

Answering complex client enquiries meant searching manually for precedents and drafting replies that held to the organization's established tone and communication standards. The manual effort per enquiry was substantial.

Solution

A human-in-the-loop drafting system built on semantic search. The system retrieves relevant knowledge and historical precedents, then proposes a response aligned with the organization's own tone and past positions. Every output is reviewed and approved by a human operator before it leaves.

Benefits

Conservative ROI of €56,000 to €76,000 from measured time savings alone. Additional value, unquantified, came from higher response quality and automated adherence to communication standards.

Technical solution

Architecture

Semantic search pipeline coupled to a generative drafting agent, tuned for tone alignment and precedent retrieval.

Technologies

Python, LangChain orchestration, sentence-transformer embeddings, Qdrant or Elasticsearch for the precedent index, hybrid retrieval with reranking, on-premises or private-tenant LLM, FastAPI, React review interface, Langfuse tracing

Slashing Chatbot Operating Costs by 30x

altFINS.com

Redesigned routing cut chatbot operating costs thirtyfold and made it reliably trigger the right financial analytics.

#Finance
#Service - Helpdesk

Problem

The first implementation of the customer-facing chatbot routed messages poorly to the underlying analytics tools and ran up large operating costs.

Solution

A conceptual redesign of the routing logic. The new architecture intercepts the user message and directs it to the right analytical pipeline. The work also laid the groundwork for Model Context Protocol integration and for orchestrating multiple ad-hoc analytics directly inside the chat interface.

Benefits

30x reduction in operating costs against the original implementation. Production accuracy rose, with few on-the-fly fixes needed after deployment.

Technical solution

Architecture

Intent classification and routing layer that decouples the conversational interface from the analytical execution underneath.

Technologies

Python, LLM function calling and structured outputs for intent routing, FastAPI, Redis for session state, Model Context Protocol servers for the analytics tools, prompt versioning and evaluation harness

Optimizing Ad Spend via Creditworthiness Prediction

hyperia.sk

A three-variable spreadsheet was replaced by a backtested 15-feature model, lifting credit prediction accuracy 20% and sharpening ad targeting.

#Finance
#Marketing - Segmentation

Problem

Advertising spend needed to be targeted at online loan applicants likely to qualify. The legacy model was a manually derived spreadsheet built on three parameters, too coarse to direct ad budget with any precision.

Solution

A stable, interpretable machine learning model trained on historical loan applicant records. The system evaluated 60 candidate features and narrowed them algorithmically to the 15 most predictive, yielding business rules the client could act on directly, such as homeowners being more than twice as creditworthy as non-homeowners.

Benefits

Prediction accuracy up 20% over the legacy model. Identifying applicants likely to receive higher loan payouts, and therefore higher commissions, let the client concentrate ad spend on high-value leads.

Technical solution

Architecture

Tabular ML pipeline built for interpretability and rigorous backtesting. Validated on a strict rolling window, 3 months training data, a 1-month gap and a 1-month prediction horizon, across 26 evaluation iterations.

Technologies

Python, pandas, scikit-learn, XGBoost and LightGBM, recursive feature elimination, weight-of-evidence binning for interpretable rules, SHAP, rolling-window backtesting harness

Uncovering Drivers of Member Retention

ver.di

Historical membership data was analyzed to isolate the behavioral profile where churn probability drops by a factor of six.

#Nonprofit
#Sales - Retention

Problem

The organization needed a data-driven picture of member behavior in order to spot flight risks early and put retention resources where they would matter.

Solution

Tabular machine learning applied to historical membership data to identify which features and behaviors correlate most strongly with retention.

Benefits

Isolated a specific combination of features marking members six times less likely to churn, giving the client a precise target profile for engagement campaigns.

Technical solution

Architecture

Exploratory data analysis and predictive modeling focused on feature importance and interpretability.

Technologies

Python, pandas, scikit-learn, XGBoost, SHAP for feature attribution, matplotlib and seaborn for the exploratory phase, SQL over the membership database

Regional Labour Market Projections for Germany

German labour-market organization

Employment forecasting moved from national aggregates to projections per region and sector, each one explainable to the planners using it.

#Government
#Strategy - Planning

Problem

Workforce planning and resource allocation were being made against national-level aggregates, which hide the divergence between regions and sectors. Decisions about where to place staff, campaigns and funding had no forecast granular enough to act on.

Solution

Sector-specific employment projections produced per geographic unit rather than nationally, combining historical labour market data with geospatial features. Output is a forecast per region and sector, with the drivers behind each projection exposed so planners can see why a region is trending as it is.

Benefits

Planning moved from national averages to region-and-sector level, letting resources be directed at the areas where the projections diverge most from the national picture.

Technical solution

Architecture

Tabular ML pipeline with geospatial feature engineering, trained per sector and evaluated on held-out time periods, with interpretability retained so projections can be explained to non-technical stakeholders.

Technologies

Python, XGBoost and LightGBM, scikit-learn, GeoPandas, PostGIS on Postgres, SHAP for feature attribution

The Onion Framework: Stress-Testing RAG Reliability

Labour Digital Slovakia

A layered evaluation method that finds where a RAG system fails before production does, not after.

#Technology
#Technology - Testing

Problem

RAG systems demo well and fail quietly. Retrieval misses, answers drift from the source, and the failure only surfaces once a user has acted on a wrong answer in production. Teams had no systematic way to know whether a system was ready to ship.

Solution

A multi-layered evaluation methodology that tests a RAG system layer by layer, from retrieval quality inward to the grounding of the final answer, so a failure is attributed to the stage that caused it rather than reported as one aggregate score. Applied as a gate before production rather than a post-mortem after it.

Benefits

Failure modes are identified and attributed before deployment rather than in production, cutting the volume of on-the-fly fixes after go-live and giving a defensible answer to the question of whether a system is ready to ship.

Technical solution

Architecture

Layered test harness over the retrieval and generation stages, with LLM-as-a-Judge scoring plus curated ground-truth sets, run as a repeatable suite rather than a one-off assessment.

Technologies

Python, Ragas and DeepEval, LLM-as-a-Judge scoring, pytest harness, Langfuse or LangSmith for trace capture, curated ground-truth sets

GenAI Adoption Across Engineering and Delivery Roles

Labour Digital Slovakia

GenAI moved from scattered individual use to a defined approach per role, with sensitive data never leaving owned infrastructure.

#Technology
#Strategy - Transformation

Problem

Generative AI was being adopted ad hoc by individual developers, with no view of which roles it actually helped, what could not be sent to an external model, or where the organization's own knowledge was sitting undiscovered.

Solution

A structured adoption strategy covering coding, project management and testing, paired with a knowledge discovery layer over internal documentation so staff could find existing material instead of recreating it. Cloud and on-premises open models were used according to the sensitivity of the data involved.

Benefits

Adoption moved from individual experimentation to a defined approach per role, with sensitive material kept on infrastructure the organization controls.

Technical solution

Architecture

Semantic search over internal documentation, with model routing between hosted and on-premises open models depending on data sensitivity.

Technologies

Llama and Mistral open weights served via vLLM or Ollama, hosted LLM APIs for non-sensitive work, Qdrant or pgvector, LangChain orchestration, Python

Semantic Interoperability for E-Government Services

Access-eGov consortium, EU-funded

Citizens' goals were resolved into the right sequence of public services across authorities, instead of being left to work it out themselves.

#Government
#Service - Knowledge

Problem

Citizens dealing with a life event such as a relocation or a business registration had to work out for themselves which authorities were involved and in what order. Each administration published its services in its own structure, so nothing could be composed across institutional boundaries.

Solution

Public services were semantically annotated against a shared model, letting a citizen's goal be resolved into the sequence of services that satisfies it, across authorities. The same annotation approach was extended to heterogeneous document repositories, making digitized collections searchable by meaning rather than by the structure each institution happened to use.

Benefits

Services from separate authorities became composable into a single guided process, and previously siloed collections became retrievable through one semantic layer.

Technical solution

Architecture

Ontology-based service annotation with a resolution layer that assembles applicable services into a goal-driven sequence, plus semantic annotation pipelines for heterogeneous document collections.

Technologies

WSMO and WSML service modelling, OWL, Protege, Sesame RDF store, Java EE, Apache Tomcat, Apache Axis, XML

Semantic IoT Integration for Energy and Cost Performance

ebbits consortium, EU-funded

Plant-floor energy events were made legible to enterprise systems, so procurement could respond to reality rather than assumptions.

#Energy
#Facilities - Energy

Problem

Enterprise systems planned against schedules while the physical plant ran against reality. Sensor and device data sat in a separate world from ERP, so energy cost was visible only after the fact and procurement decisions ignored what was actually happening on the floor.

Solution

Device and sensor data was semantically integrated into enterprise systems, so real-world energy events became something business applications could act on. Context-aware processing let procurement respond to actual energy conditions rather than to static assumptions.

Benefits

Energy and cost performance became visible and actionable inside the enterprise systems already in use, closing the gap between plant reality and business decisions.

Technical solution

Architecture

Semantic middleware layer bridging device networks and enterprise applications, with event processing that translates physical events into business-level context.

Technologies

LinkSmart (Hydra) semantic middleware, OSGi Equinox, Java, RDF and OWL device ontologies, CoAP and MQTT device protocols, Esper CEP, SOAP and REST connectors to ERP

Data-Driven Resource Efficiency for Process Industries

MONSOON consortium, EU-funded

Process operators and data scientists built one shared model of the plant, turning operator knowledge into something the models could actually use.

#Energy
#Supply - Maintenance

Problem

Process plants generate continuous data but the knowledge needed to interpret it sits with domain experts, not data scientists. Energy and raw material losses were understood qualitatively by the people running the process and not represented anywhere a model could use.

Solution

Data-driven tools for the efficient use and re-use of energy and raw resources, built on a shared domain model. A collaborative web-based modelling environment let process experts and data scientists build that model together, so what the operators knew ended up encoded rather than lost in translation.

Benefits

Domain knowledge was captured in a form models could consume, and resource efficiency decisions gained a data basis instead of resting on operator intuition alone.

Technical solution

Architecture

Shared industrial domain model built collaboratively, feeding predictive models over process time-series data.

Technologies

Apache NiFi, Apache Kafka with MQTT, Hadoop HDFS, Apache Spark, DL4J with Keras, Apache Zeppelin, Grafana, PFA and PMML model interchange, JSON-LD semantic models aligned to CRISP-DM and ISO 22400

Comorbidity Decision Support for Specialists

Healthcare research partner

Everything clinically relevant across a patient's conditions, in one view, organized by what interacts rather than by which system stored it.

#Healthcare
#Service - Knowledge

Problem

Chronic disease management breaks down when a patient has several conditions at once. The relevant history is scattered across systems and specialties, and the specialist has minutes to find what matters. The harder problem is not collecting the information but presenting it so the clinically significant interaction is visible.

Solution

Patient information relevant to the specialist's own question is collected across conditions and presented in one view, organized around the interactions between comorbidities rather than as separate per-condition records.

Benefits

The specialist sees the clinically relevant picture across all of a patient's conditions in one place, rather than reconstructing it from separate records under time pressure.

Technical solution

Architecture

Aggregation across clinical sources into a unified patient model, with a presentation layer organized by clinical relevance rather than by source system.

Technologies

HL7 v2/v3 and later FHIR interfaces, SNOMED CT and ICD-10 vocabularies, OWL patient and comorbidity models, Java EE, D3.js visualization, SQL clinical data store

Public Procurement Data Analytics

Transparex.sk

Public procurement notices were turned from nested XML into a dataset where patterns across authorities and suppliers are visible at scale.

#Government
#Procurement - Tendering

Problem

Public procurement is published as open data but not as usable data. Notices arrive in deeply nested and inconsistent XML structures, which meant that questions about who wins what, how often, and at what price could not be answered without substantial manual work per question.

Solution

Complex procurement structures were parsed and normalized into a queryable model, turning published notices into a dataset that supports analysis across contracting authorities, suppliers and time.

Benefits

Procurement questions became answerable by query rather than by manual inspection, making patterns across authorities and suppliers visible at scale.

Technical solution

Architecture

Extraction and normalization pipeline over heterogeneous procurement XML into a relational model built for analytical queries.

Technologies

Python, XML, XSLT, SQL

Customer Support RAG Chatbot

SME client

Support answers come straight from the company's own documentation, day or night, instead of being retyped by staff.

#Legal
#Service - Helpdesk

Problem

Support answers already existed in the company's documentation, and customers still asked the team. Staff spent their day retyping answers that were written down somewhere, and anyone asking outside working hours waited until morning for a reply that was already in a file.

Solution

A context-aware chatbot answering from the company's own documents rather than from general knowledge, so responses reflect the current documentation and stay inside what the company has actually published. Questions it cannot answer from the source material are handed to a person instead of guessed at.

Benefits

Routine repeat questions are answered from the documentation without a person in the loop, and answers become available outside working hours. Estimated: the majority of first-line questions deflected, with a working prototype delivered in weeks rather than months.

Technical solution

Architecture

Documentation is chunked and embedded into a vector store, retrieved per question and passed to the model with the source context, served through an API behind the chat interface.

Technologies

LangChain, FastAPI, ChromaDB, OpenAI, Python

B2B Lead Generation Agent

Internal product, offered to B2B clients

Prospect lists built and qualified unattended, with the same criteria applied to the two-hundredth company as the first.

#Technology
#Sales - Prospecting

Problem

Building a prospect list means opening business listings one by one, visiting each website, and judging whether the company is worth approaching. It is a full day of work for a list that goes stale in a month, and the judgement applied on result number two hundred is not the judgement applied on result number two.

Solution

An autonomous agent that finds candidate companies through Google Maps, visits each website, and classifies whether it fits the target profile, applying the same criteria to every prospect. The output is a qualified list rather than raw search results, stored so it can be refreshed rather than rebuilt.

Benefits

Lead research runs unattended instead of consuming a person's day, and qualification criteria are applied consistently across the whole list rather than drifting as the researcher tires. Estimated: several hours of manual research replaced per list, with qualification consistency being the larger gain over speed.

Technical solution

Architecture

Discovery through the mapping source, browser automation to visit and read each candidate site, model-based classification against the target profile, results persisted to a relational store and reviewed through a web interface.

Technologies

Playwright, FastAPI, PostgreSQL, React, Python, LLM-based classification

AI-Assisted Slovak Tax Return Filing

Student users

Tax documents read automatically into the right fields of the Slovak return, validated and explained before anything is submitted.

#Finance
#Finance - Taxation

Problem

A student filing a Slovak tax return for the first time faces a form written for accountants. The information needed is sitting on documents they already hold, but transferring it means knowing which figure belongs in which field, and a mistake is only discovered when the return is rejected.

Solution

Source documents are read automatically and their figures placed into the correct fields of the Slovak tax form, with validation catching inconsistencies before submission and plain-language explanations of what each field is asking for, so the user understands the return rather than just completes it.

Benefits

First-time filers complete a return from their own documents without an accountant, with errors caught before submission rather than after rejection. Estimated: filing time reduced from hours of form-reading to minutes of review, with validation preventing the common field-placement errors.

Technical solution

Architecture

Document ingestion and field extraction through a document intelligence service, mapped onto the tax form schema with a validation layer over the extracted values and an explanation layer surfacing what each field requires.

Technologies

Python, Azure Document Intelligence, TypeScript

Multi-Agent Router for Customer Email at Scale

GymBeam

Thirteen thousand customer emails a month across thirteen markets, routed by intent to seven specialized agents that resolve them end-to-end.

#Retail
#Service - Ticketing

Problem

Around 13,000 customer emails a month arrive across 13 European markets, in as many languages, covering everything from where-is-my-order to returns to product questions. A single general-purpose assistant handles all of them adequately and none of them well, and a human triage layer scales only by hiring.

Solution

An LLM router classifies the intent of each incoming email and dispatches it to one of seven independent agents, each built and deployed for a single category of request. Specialization is the point: each agent can be evaluated, tuned and improved against its own intent without regressing the others, and a new intent means a new agent rather than a longer prompt.

Benefits

Roughly 13,000 emails a month routed and resolved end-to-end across 13 markets, with seven agents deployed independently so a change to one intent carries no risk to the rest. Handling capacity no longer scales with headcount.

Technical solution

Architecture

Intent classification at the entry point dispatching to seven independently deployed agents on Cloud Run, each scoped to one request category, scaling per intent rather than as one monolith.

Technologies

Python, LangChain, Google ADK, Gemini, Google Cloud Run, FastAPI, BigQuery

Tool-Calling Returns Agent

GymBeam

Returns validated against the live order, labelled and answered in the customer's language, without anyone touching the ticket.

#Retail
#Supply - Returns

Problem

A return request is not a question, it is a transaction. Someone has to find the order, confirm it qualifies, generate a label and reply in the customer's language. Across 13 markets this is the highest-volume repetitive task in customer care, and every step of it already exists as an API call.

Solution

An agent that resolves the request itself rather than drafting a reply about it: the order is validated through the Magento API, a return label is issued, and the customer is answered in their own language. The agent calls the systems of record directly, so what the customer is told matches what the platform actually did.

Benefits

Return requests close without human involvement across 13 markets and languages, with the reply grounded in a real order lookup rather than a generated answer. Estimated: the largest single category of repetitive customer-care work removed, with response time falling from hours to the length of the API round trip.

Technical solution

Architecture

Tool-calling agent with the order platform exposed as callable functions, validating eligibility against live order data before issuing a label, with the response generated in the customer's market language.

Technologies

Gemini, LangChain, Magento REST API, Python, FastAPI, Google Cloud Run

"Jim", a Conversational Shopping Assistant

GymBeam

A shopping assistant that recommends from the real catalogue, embedded in the storefront where the buying decision happens.

#Retail
#Sales - Upselling

Problem

Customers arrive knowing what they want to achieve, not which SKU achieves it. Search matches words against a catalogue of thousands of products, so a shopper who cannot name what they need either filters blindly or leaves. The product knowledge that would answer them exists, but only in the catalogue data.

Solution

A conversational assistant, live on the e-shop as a JavaScript widget, that answers from the product catalogue itself through retrieval rather than from general model knowledge. Recommendations refer to products that actually exist and are actually in the range.

Benefits

Shoppers who cannot name what they need get a recommendation from the real catalogue instead of abandoning search, on the storefront rather than in a separate channel. Estimated: measurable lift in assisted conversion and reduced exit rate from search, subject to A/B measurement.

Technical solution

Architecture

RAG over the product catalogue behind a FastAPI service, embedded into the storefront as a JavaScript widget so the assistant sits in the purchase path rather than beside it.

Technologies

FastAPI, Google ADK, Gemini, vector database, JavaScript widget, Python

Production Guardrails and LLM Observability

GymBeam

Every agent action traced, costed and dry-run against historical tickets, so automation expands on evidence rather than hope.

#Retail
#Technology - Monitoring

Problem

An agent answering real customers is one bad answer away from a complaint, and the failure is invisible unless someone is watching. Teams running LLMs in production typically cannot say what an agent did on a given ticket, what it cost, or whether last week's prompt change made things worse.

Solution

A production control layer around the agents: low-confidence cases hand off to a human rather than guessing, every execution is traced end to end, token cost is tracked in the warehouse, and changes are dry-run against historical tickets before release, so a regression is caught against real past traffic rather than in front of a customer.

Benefits

Agent behaviour, cost and failure modes became observable rather than assumed, and changes are validated against historical tickets before they reach live traffic. Estimated: the main enabler of the wider rollout, since each additional automated intent depends on being able to prove it behaves.

Technical solution

Architecture

Confidence thresholds gating human handoff, full execution tracing per request, token-cost accounting into BigQuery, dry-run evaluation harness replaying historical tickets, with SQL transformations feeding monitoring dashboards.

Technologies

BigQuery, SQL, Tableau, LangSmith-style execution tracing, Python, prompt and eval versioning

GraphRAG over Slovak Legislation

Technical University of Košice

Legal answers retrieved through the reference structure of the law itself, every one traceable back to the provision it came from.

#Legal
#Legal - Compliance

Problem

Legislation is a graph pretending to be a library. Provisions amend, reference and depend on each other, so the passage that answers a question is often not the passage that matches its words. Vector search alone retrieves what sounds relevant and misses the cross-reference that changes the answer, and in a legal context an answer without a citation is unusable.

Solution

A GraphRAG system combining a graph representation of the legislation with vector search, so retrieval follows the actual structure of references between provisions as well as semantic similarity. Every answer carries the citations it rests on, making it checkable against the source text.

Benefits

Legal answers that follow the reference structure of the legislation rather than surface wording, each one citation-backed and verifiable against the original provision. Estimated: accuracy gains over pure vector retrieval concentrated exactly where it matters, on questions whose answer depends on a cross-reference.

Technical solution

Architecture

Legislation modelled as a graph of provisions and their references in Neo4j, combined with vector retrieval over provision text, with citations carried through to the generated answer.

Technologies

Neo4j, vector database, Python, LangChain, LLM generation with citation grounding

RAG Pipelines for Private Domain Documents

Siemens Healthineers

Private engineering documentation made answerable by question, with nothing leaving the organization's own infrastructure.

#Healthcare
#Service - Knowledge

Problem

Engineering and regulatory knowledge in medical technology sits in private document sets that a general-purpose model has never seen and must never be trained on. The knowledge is written down and still effectively unsearchable, because finding it requires knowing which document it is in.

Solution

Retrieval-augmented pipelines letting language models work over domain-specific and private documents, so answers come from the organization's own material with the documents staying under its control rather than being absorbed into a model.

Benefits

Internal document sets became answerable by question rather than by knowing where to look, without private material leaving the organization's control. Estimated: substantial reduction in time spent locating information across internal documentation.

Technical solution

Architecture

Document ingestion and chunking into a vector index over private corpora, retrieval feeding a generation layer, with the source documents remaining inside controlled infrastructure.

Technologies

Python, LangChain, vector databases, embedding models, RAG orchestration

Siml.ai: A Platform for Learnable Simulators

DimensionLab

Simulations that took weeks run in minutes, because the physics is learned once and then queried rather than solved again every time.

#Manufacturing
#Product - Engineering

Problem

Designing a fusion reactor, a wind farm, a carbon capture system or a hydrogen battery means running the same expensive simulation hundreds of times with slightly different parameters. Each run is computationally costly and each one is thrown away, so the experimental phase of development takes weeks or months and the number of designs anyone can explore is set by compute budget rather than by engineering judgement.

Solution

A platform where engineers build, train and optimize AI surrogate models that approximate the physics rather than solving it from scratch each time. The surrogate is trained once and then queried in real time, so the cost sits in training rather than in every subsequent run, and exploring a design space becomes interactive instead of a queue of overnight jobs. Physics-informed training means the model is constrained by the governing equations rather than only by data, which matters where experimental data is scarce or expensive.

Benefits

One customer achieved 96% cost and time savings on AI-driven simulations built on Navier-Stokes equations and physics-informed neural networks. Experimental phases that ran to weeks or months compress to hours, and in some cases minutes. The platform passed 500 users during early release.

Technical solution

Architecture

Physics-informed neural networks and neural operators trained against governing PDEs and, where available, experimental data, producing hybrid physics-plus-data surrogate models. Trained once, then deployed for repeated real-time inference.

Technologies

NVIDIA Modulus and PhysicsNeMo, PyTorch, physics-informed neural networks, Fourier and graph neural operators, CUDA on NVIDIA GPUs, Docker, cloud-agnostic deployment

SITE: Simulator Inference and Training Environment

DimensionLab

The whole physics-ML stack in one container, so engineers start from the simulation question rather than from infrastructure.

#Manufacturing
#Technology - Infrastructure

Problem

Physics-ML sits between two worlds that do not share tooling. Getting a physics-informed model trained means assembling GPU infrastructure, geometry handling, a training loop, rendering and hardware monitoring, and most of that assembly work has nothing to do with the engineering question being asked. It is the reason the technique stays with specialists.

Solution

A packaged training and inference environment optimized for NVIDIA GPUs, with the physics-ML framework, tuned training and inference pipelines, high-performance rendering and real-time hardware monitoring integrated into one containerized setup that runs on any cloud. The engineer works on the simulation rather than on the stack underneath it.

Benefits

Physics-ML training becomes accessible without assembling the infrastructure first, and the cloud-agnostic containerized setup avoids tying customers to one provider. Estimated: the setup work ahead of a first trained model reduced from an infrastructure project to a container launch.

Technical solution

Architecture

Dockerized, cloud-agnostic environment integrating the physics-ML framework with training and inference pipelines optimized in-house, plus geometry handling, rendering and live hardware utilization monitoring.

Technologies

NVIDIA Modulus, Docker, CUDA, GPU cluster orchestration, geometry processing including CSG and tessellated STL, real-time rendering

Neural Operators for Complex-Geometry CFD

DimensionLab

Fluid dynamics surrogates that handle the irregular geometry of real parts, not just the clean shapes that suit the maths.

#Manufacturing
#Strategy - Innovation

Problem

Computational fluid dynamics on real geometries is where surrogate models usually break. Irregular meshes do not suit architectures built for regular grids, and the processing cost of traditional methods on complex shapes is exactly what makes design iteration slow. A surrogate that only works on simple geometry solves the easy half of the problem.

Solution

Geometry-Informed Neural Operators, combining a graph neural operator layer that handles irregular grids with a Fourier neural operator for the downstream processing. The result adapts to complex geometries rather than requiring them to be simplified, which is what makes the approach usable on real engineering parts.

Benefits

Faster and more adaptable CFD prediction on the complex geometries where conventional methods are most expensive, extending surrogate modelling from demonstration shapes to real engineering parts. Estimated: the enabling capability behind customer CFD deployments rather than a standalone result.

Technical solution

Architecture

Graph neural operator layer ingesting irregular meshes, feeding a Fourier neural operator for spectral processing, trained as a geometry-conditioned surrogate over CFD workloads.

Technologies

Graph Neural Operators, Fourier Neural Operators, PyTorch, NVIDIA Modulus, CUDA

Agentic Layer for AI Engineering Workflows

DimensionLab

Agents assisting simulation setup and analysis, with evaluations and guardrails deciding where they are allowed to act alone.

#Manufacturing
#Technology - Development

Problem

Setting up a simulation is itself expert work: choosing the formulation, the boundary conditions, the training strategy. That expertise gates who can use the platform, and it is the kind of structured, repeatable judgement that language models can now assist with, provided the assistance can be trusted in a domain where a plausible-looking wrong answer is worse than no answer.

Solution

An agentic layer over the engineering workflow, orchestrating language models across the setup and analysis steps, with evaluation harnesses and guardrails constraining what the agents are permitted to conclude and where a human must confirm. Model choice is kept open across providers rather than committed to one vendor.

Benefits

Simulation setup and analysis assisted by agents whose behaviour is measured rather than assumed, widening who can operate the platform. Estimated throughout: this entry is reconstructed from a technology list, not from project documentation.

Technical solution

Architecture

Multi-agent orchestration across the engineering workflow, with evaluation suites and guardrails gating agent output and provider-agnostic model routing.

Technologies

Paperclip agent orchestration, Anthropic Claude, OpenAI and open Chinese model families, evaluation harnesses, guardrails, Python

Agentic AI Platform for an Insurance Group

Insurance group

An insurance agent platform where new workflows are declared rather than coded, running as background tasks triggered by the data itself.

#Insurance
#Legal - Insurance

Problem

Insurance work is a long tail of judgement-heavy tasks sitting on top of records that change constantly. Automating any one of them individually produces a tool that cannot be extended, and enterprise procurement will not accept a system that has to be rebuilt each time a new process is added. The requirement is a platform that absorbs new workflows rather than a set of point solutions.

Solution

An agentic platform combining chat-based agents with declarative agentic workflows. Most workflows run as background tasks triggered by events rather than by users: a new record arrives in the database that needs evaluating, and the workflow starts itself. Orchestration is declarative rather than code-based, so adding or changing a workflow is a configuration change instead of a system rebuild, which is what makes the platform absorb new use cases rather than accumulate bespoke code.

Benefits

New workflows onboard without rebuilding the system, and event-triggered background execution means throughput is not bound by user interaction. Estimated: three business use cases are being migrated onto the platform progressively, indicating it is being treated as the target architecture rather than a pilot.

Technical solution

Architecture

Declarative multi-agent orchestration on Microsoft Agent Framework, with durable workflow execution handling long-running and resumable processes, vector and relational stores behind retrieval, and a deployment model spanning Azure and on-premises inference so data location can be controlled per workload.

Technologies

C# and .NET, Microsoft Agent Framework (MAF), Temporal, Qdrant, PostgreSQL, MS SQL, S3-compatible storage, Azure deployment, on-premises Mistral, DeepSeek, Qwen and open-source models, embedding and vision models

Chat-Based LLM Agents for Critical Infrastructure

Slovak and Czech organizations in critical infrastructure, including national security and audit authorities

Retrieval chat agents deployed inside critical infrastructure and security authorities, built before the standard frameworks existed.

#Government
#Service - Knowledge

Problem

Organizations handling classified or critical-infrastructure information had the same knowledge-access problem as everyone else and none of the available options. Sending queries to a commercial API was not permissible, and at the time the work started the tooling everyone now takes for granted did not exist.

Solution

LLM chatbots with retrieval over internal document sets, built from first principles because the frameworks that would later standardize this work had not been released. The systems grew by accretion, retrieval first, then progressively more capable agentic behaviour, with deployment moving to Azure as soon as the models became available there so that data location met the organizations' requirements.

Benefits

Internal knowledge became queryable in organizations where commercial AI services were not an option. Estimated: the early-mover position produced hard-won production experience in exactly the environments where AI deployment is hardest to get approved.

Technical solution

Architecture

Retrieval-augmented chat over internal corpora, hand-built orchestration predating the standard frameworks, later migrated to Azure-hosted models for data residency compliance.

Technologies

OpenAI services initially, later Azure OpenAI for data location, custom retrieval and orchestration code, vector search, .NET

Slovak Foundation Models

SlovakNLP, with the Institute of Artificial Intelligence TUKE and the Slovak Academy of Sciences

Open-weight Slovak models that can be self-hosted and adapted, closing the gap that made Slovak an API-only language.

#Research
#Product - Engineering

Problem

Slovak is a low-resource language. Open-weight models handle it badly, switching mid-sentence into other languages, looping, and pulling vocabulary from neighbouring Slavic languages. The commercial models that do handle Slovak well are closed, which rules them out wherever the data cannot leave the organization or the model has to be adapted to a domain.

Solution

A Slovak baseline model built by full-parameter fine-tuning of Mistral 7B on a large Slovak web corpus, followed by an instruction-tuned 14B model on the Qwen architecture. The result is an open-weight model that can be self-hosted, inspected and further adapted, rather than accessed through someone else's API.

Benefits

Grammatically correct and contextually coherent Slovak, with the code-switching, repetition and lexical interference of the base model eliminated. Available as open weights, so organizations handling regulated or confidential Slovak text can run them on their own infrastructure.

Technical solution

Architecture

Full-parameter fine-tuning of open-weight base models on a large Slovak corpus, followed by instruction tuning on curated Slovak sources, trained on national HPC allocation.

Technologies

Mistral 7B and Qwen 14B architectures, full-parameter fine-tuning, Araneum Slovacum corpus, Leonardo supercomputer, Hugging Face distribution

Predicting the Outcome of Judicial Appeals

SlovakNLP

Appeal odds estimated from the text of the decision itself, trained on a corpus of millions of Slovak court rulings.

#Legal
#Legal - Litigation

Problem

Deciding whether to appeal is a judgement made with almost no data behind it. Slovak court decisions are digitized in the millions but written as unstructured prose, so nobody can say what the realistic odds of reversal look like for a case of a given shape.

Solution

Court decisions are segmented into their logical parts, then domain-adapted language models are fine-tuned on Slovak legal text and applied to predict whether an appeal will succeed. Candidate models were selected by benchmarking a field of nine against generation metrics before adaptation, and the prediction itself was approached three ways, zero-shot, few-shot and with a dedicated classification layer, so the cheapest adequate method could be identified rather than assumed.

Benefits

Appeal prospects become estimable from the case text itself rather than from intuition, on a corpus drawn from millions of digitized Slovak decisions.

Technical solution

Architecture

Rule-based structural segmentation of decision texts, legal-domain adaptation of Slovak base models, then appeal-outcome prediction evaluated across zero-shot, few-shot and classification-head approaches.

Technologies

Qwen3-8B and Mistral-SK-7B, domain-adaptive fine-tuning, classification heads, Perplexity, BLEU and ROUGE benchmarking

Semantic Search and RAG Across 70 Million Documents

Dr. Evidence

Semantic search and RAG engineered to hold their accuracy at 70 million documents, not just at demo scale.

#Pharmacy
#Service - Knowledge

Problem

Medical and regulatory intelligence means searching the entire published literature, and RAG architectures that work on a corporate wiki fall apart at that scale. Retrieval latency, index cost and recall all degrade, and in a regulatory context an answer that misses the relevant paper is worse than no answer.

Solution

Semantic search and retrieval-augmented generation scaled to a corpus of 70 million documents, engineered so that retrieval quality holds at production scale rather than degrading as the index grows.

Benefits

Literature search across a corpus of tens of millions of documents became a single semantic query instead of keyword searches across siloed databases followed by manual synthesis.

Technical solution

Architecture

Distributed vector and lexical indexing over a corpus of tens of millions of documents, with hybrid retrieval feeding a generation layer built for regulatory-grade traceability back to source.

Technologies

Elasticsearch and Solr, Apache Spark, transformer embedding models, approximate nearest-neighbour vector indexing (FAISS/HNSW), cross-encoder reranking, Kubernetes, Python

Information Extraction Models for Medical and Pharmaceutical Text

Dr. Evidence

Extraction models tuned for biomedical text, turning published literature into structured records that can actually be compared.

#Pharmacy
#Data - Engineering

Problem

The value in a clinical paper is in its structured detail, the population, the intervention, the endpoints, the effect sizes. General-purpose models extract this unreliably in the biomedical domain, where the same concept appears under a dozen names and a misread number is a clinical error rather than a typo.

Solution

Instruction-fine-tuned models built specifically for extraction in the medical and pharmaceutical domains, trained to return structured fields rather than prose, so downstream comparison and analysis operate on data instead of on text.

Benefits

Structured extraction that holds up in a domain where general models are unreliable, turning published literature into comparable structured records at scale.

Technical solution

Architecture

Domain-specific instruction fine-tuning for structured extraction, with schema-constrained output over biomedical source documents.

Technologies

Hugging Face Transformers, PEFT/LoRA fine-tuning, PyTorch, constrained JSON-schema decoding, UMLS, MeSH and SNOMED CT vocabularies, spaCy/scispaCy, Python

Agentic Platform for the Biopharmaceutical Domain

Dr. Evidence

More than 25 specialized agents across three workflow types, each narrow enough that its output can be checked rather than trusted.

#Pharmacy
#Legal - Compliance

Problem

Regulatory and labelling work is a chain of distinct tasks, each needing different sources, different judgement and different output. A single general assistant handles all of them mediocrely, and mediocre is not a usable standard when the output supports a regulatory submission.

Solution

An agentic platform orchestrating more than 25 specialized agents across three workflow types, each agent scoped to a task it can be evaluated on individually, with the orchestration layer deciding which agents a given piece of work requires.

Benefits

Regulatory and labelling workflows that previously required specialist manual effort at every step run as orchestrated agent workflows, with each agent narrow enough for its output to be verified rather than trusted wholesale.

Technical solution

Architecture

Multi-agent orchestration across three workflow families, with 25+ task-specialized agents over a curated document corpus, scoped so that each agent's output is individually evaluable.

Technologies

LangChain and LangGraph, vector store retrieval, fine-tuned open-weight and hosted LLMs, FastAPI, Kubernetes, LangSmith/Langfuse tracing, Python

Big Data Platform for Predictive Control and Maintenance

MONSOON consortium, EU SPIRE-723650

Plant data processed fast enough to act on, with the analytics described semantically so they move between sites instead of being rebuilt.

#Energy
#Supply - Maintenance

Problem

Process plants generate continuous high-volume sensor data, and the value in it, predicting a failure or correcting a drifting process before it produces scrap, is only realized if the data can be processed fast enough to act on. Batch analysis after the shift tells you what you already lost.

Solution

A big data platform supporting predictive control and predictive maintenance over process data, paired with a semantic model that represents the data-analytics processes themselves, so an analytical pipeline is described in a form that can be reused across plants rather than rebuilt for each one.

Benefits

Predictive control and maintenance became possible on live process data, and the analytics themselves became portable between sites through a shared semantic description.

Technical solution

Architecture

Distributed batch and stream processing over plant sensor data, with a semantic layer describing analytical processes for reuse across deployments.

Technologies

Apache NiFi (plant-side integration), MQTT, Apache Kafka, Apache Hadoop HDFS and YARN, Apache Spark (batch, streaming and SQL), DL4J with Keras for GPU/CPU training, Hadrian scoring engine for PFA, Grafana, Apache Zeppelin, Apache Knox, Apache Ambari, OSIsoft PI historian connector. Semantic framework built on CRISP-DM, PMML, PFA, ISO 22400 KPIs and JSON-LD.

Collaborative Policy Modelling Platform

InterSoft a.s., partner in the OCOPOMO consortium, EU FP7 248128

Stakeholders write the scenarios, the simulation traces back to them, and policy models stop being a black box to everyone outside the modelling team.

#Government
#Strategy - Development

Problem

Policy models are built by modellers and the evidence behind them lives with stakeholders, in narrative form. The gap between the two is where trust in a policy simulation is lost: nobody outside the modelling team can trace why the model behaves as it does or which stakeholder input produced a given assumption.

Solution

An online collaboration and content management platform where stakeholders build narrative scenarios that are then transformed through conceptual models into formal policy simulations, with traceability maintained from each element of the model back to the evidence that produced it.

Benefits

Policy simulations became traceable back to stakeholder evidence rather than opaque, which is the condition for non-modellers to trust and use them. Piloted on EU structural funds and on regional renewable energy policy.

Technical solution

Architecture

Enterprise content management backbone supporting collaborative scenario authoring, with transformation into conceptual and then formal policy models, preserving traceability links throughout.

Technologies

Alfresco CMS, J2EE, agent-based simulation integration, semantic annotation

Semantic Service Bus for Networked Enterprises

SPIKE consortium, EU FP7 217098

Services from different companies discovered and matched semantically, so setting up a business alliance stops being a bespoke integration project.

#Technology
#Technology - Integration

Problem

Short-term business alliances need to share services across company boundaries, but every partner's services are described differently. Wiring each new alliance together by hand costs more than the collaboration is worth, which puts this kind of partnership out of reach for smaller firms.

Solution

Semantic extensions to an enterprise service bus, so services from different organizations are annotated against a shared model and can be discovered and matched to workflow tasks automatically, with data mediated between them rather than manually mapped.

Benefits

Cross-company service integration shifted from bespoke wiring per alliance to semantic matching, bringing the setup cost of a business alliance within reach of SMEs.

Technical solution

Architecture

Semantically enriched service bus with annotation-driven service discovery and data mediation across organizational boundaries, covering web services, cloud services and human tasks.

Technologies

SA-WSDL, Apache ServiceMix ESB, Apache CXF, OWL and WSMO ontologies, Liferay portal with portlets (JSR-168/286), OSGi, Java EE, XML SOAP and REST connectors

Semantic Platform for E-Government Services

Access-eGov consortium, EU FP6 IST-27020

A citizen states the goal and the platform assembles the right services from the right authorities, in the right order.

#Government
#Service - Knowledge

Problem

A citizen dealing with a life event faces a set of authorities that do not present themselves as a set. Each publishes its services in its own structure, so working out which apply and in what order is left to the citizen, and traditional counter services and electronic ones do not connect at all.

Solution

A platform for semantic integration of traditional and electronic government services, where services are modelled ontologically and a resolution component assembles the ones applicable to a citizen's goal into an executable sequence across institutional boundaries.

Benefits

Services from separate authorities, electronic and counter-based alike, became composable into one guided process driven by what the citizen is trying to achieve.

Technical solution

Architecture

Ontology-based modelling of government services with a core execution component that resolves a citizen goal into an ordered set of applicable services spanning authorities.

Technologies

WSMO and WSML service modelling, OWL, Protege, Sesame RDF store, Java EE, Apache Tomcat, Apache Axis web services

Object/RDF Persistence and Hybrid Semantic Search

KP-Lab consortium, EU FP6 IST-27490

Structured semantic queries and ordinary keyword search answered by one service, over an object API developers could actually use.

#Research
#Technology - Infrastructure

Problem

Knowledge-practice environments need to store richly interlinked objects and let people find them, and the two requirements pull apart. Semantic stores answer structured queries but not the vague ones people actually type; full-text search does the opposite and ignores the structure entirely.

Solution

An object/RDF persistence and query API giving applications a normal programming model over a semantic store, combined with a search service that answers semantic and free-term queries together rather than forcing a choice between them.

Benefits

Applications gained structured semantic querying without giving up ordinary keyword search, and developers worked against a familiar object API instead of writing RDF queries by hand.

Technical solution

Architecture

Object-to-RDF mapping layer with a query API over a semantic repository, and an integrated retrieval service combining semantic and free-term search paths.

Technologies

Sesame/OpenRDF repository, SPARQL, RDF and OWL, Java object-RDF mapping API, Apache Lucene, Apache Tomcat, Java EE

Trust Mechanisms and Transaction Analytics for an SME E-Marketplace

SEAMLESS consortium, EU FP6 IST-26476

Small firms got a basis for trusting cross-border counterparties, built on analytics over what actually happened on the platform.

#Retail
#Legal - Risk

Problem

Cross-border electronic marketplaces fail for small enterprises on trust rather than on technology. A firm will not transact with an unknown counterparty in another country without some basis for judging them, and marketplace operators had no analytical view of what was actually happening across transactions.

Solution

Analysis of trust-building mechanisms for an electronic marketplace and design of the operational scenarios to implement them, followed by a data warehouse with OLAP and data mining services giving the operator an analytical and rating view over platform transactions.

Benefits

Trust mechanisms moved from principle to implementable operational scenarios, and marketplace transactions became analyzable, supplying the evidence base for participant ratings.

Technical solution

Architecture

Data warehouse over marketplace transactions with OLAP cubes and data mining services feeding analytical and rating functions, alongside designed trust-building operational flows.

Technologies

JDM (Java Data Mining, JSR-73) and JOLAP APIs, Mondrian OLAP with Pentaho reporting, MySQL data warehouse, Weka and JBowl for mining, J2EE, SQL

Open-Source Text Mining Library and Analytical Platform

Technical University of Košice

An open-source Java library that made text mining a matter of composing components instead of rebuilding the pipeline every project.

#Technology
#Technology - Development

Problem

Text mining work was repeatedly rebuilt from scratch, every project reimplementing preprocessing, classification, clustering and evaluation before it could get to its actual question. Nothing modular and extensible existed for Java, and later the same code could not scale to larger corpora or run distributed.

Solution

An open-source Java library covering the whole text mining process, from corpus construction and preprocessing through classification and clustering to model evaluation, founded and maintained as a public project. It was later extended with a task-based execution engine for distributed processing and a graphical console, making it usable by people without programming skills.

Benefits

Text mining became a matter of composing an existing library rather than rebuilding infrastructure per project. The library was adopted across research projects and teaching, and the distributed engine extended it to corpora that no longer fit a single machine.

Technical solution

Architecture

Modular Java API spanning the text mining pipeline, with a task-based middleware execution engine for multi-threaded and distributed processing, and a web console over the library.

Technologies

Java SE and EE, Apache Lucene, Weka-compatible model interfaces, task-based middleware execution engine, Apache Hadoop for distributed runs, Apache Tomcat web console, distributed via SourceForge

Text Categorization for Direct Democratic Participation

WEBOCRACY consortium, EU IST-1999-20364

Citizen input classified and retrievable by subject, so participation at scale produces something an administration can actually answer.

#Government
#Administration - Correspondence

Problem

Opening public administration to direct citizen participation produces volume, and volume without organization is noise. Contributions arrive as free text on every subject at once, and the administration cannot route or answer what it cannot categorize.

Solution

A collaboration platform built on semantic web technologies for citizen participation in democratic processes, with text categorization and information retrieval so that contributions are classified by subject automatically and previous material on the same subject is findable.

Benefits

Citizen contributions became routable and searchable by subject rather than accumulating as an undifferentiated stream, which is what makes participation at scale answerable.

Technical solution

Architecture

Semantic web collaboration platform with automatic text categorization and retrieval services over citizen contributions.

Technologies

Webocrat system on Java and J2EE, RDF and RDF Schema, Apache Lucene retrieval, JBowl text categorization, Apache Tomcat, relational backend over SQL

Data Mining over Integrated GIS and Data Warehouse

GOAL consortium, INCO-COPERNICUS 977091

Spatial and analytical questions answered in one system, from regional water demand forecasting to scenario discovery for tourism operators.

#Utilities
#Supply - Planning

Problem

Geographic systems and analytical data warehouses were separate worlds. Questions with both a spatial and an analytical dimension, how consumption varies across a distribution network, which regions support which service scenarios, could not be asked in either system alone.

Solution

A data mining application integrated directly into a GIS over a data warehouse, applied to predicting water consumption across a regional distribution system and to identifying optimal operating scenarios for travel agencies.

Benefits

Spatial and analytical questions became answerable in one system, demonstrated on regional water consumption forecasting and on scenario discovery for tourism operators.

Technical solution

Architecture

Data mining components integrated with a GIS front end over OLAP services on a data warehouse, supporting spatially resolved prediction and scenario analysis.

Technologies

C/C++, MS SQL OLAP Services, ADO, GIS integration

Vowel Detection in Noisy Speech

UNIZA, University of Žilina

Treating noisy speech as images rather than sequences raised vowel detection accuracy by up to 16 points over published alternatives.

#Research
#Strategy - Innovation

Problem

Speech recognition degrades sharply once background noise enters the signal, and vowel segments are where that degradation does most damage. Existing approaches based on attention networks were leaving accuracy on the table under realistic noise conditions.

Solution

Audio is transformed into a spectrogram and individual segments are classified with convolutional neural networks, treating vowel detection as an image classification problem rather than a sequence modelling one.

Benefits

Detection accuracy 6 to 16 percentage points higher than comparable research using attention-based networks. Published and peer-reviewed: https://doi.org/10.1016/j.trpro.2021.07.112

Technical solution

Architecture

Spectrogram transformation of the audio signal followed by segment-level CNN classification, GPU-accelerated for training throughput.

Technologies

Python, CuPy, convolutional neural networks

Pairwise Weighted Classifier Ensembles

UNIZA, University of Žilina

A new weighting scheme for combining pairwise classifiers beat the best existing combination method on both CIFAR-100 and ImageNet.

#Research
#Strategy - Innovation

Problem

Multi-class classification can be decomposed into pairwise classifiers, but combining their outputs back into a single decision is where accuracy leaks away. Standard combination schemes were failing to extract the full value of strong individual models.

Solution

A new pairwise weighted ensemble method that combines the outputs of deep convolutional networks and vision transformers through tensor operations, weighting each pairwise decision by its reliability rather than treating all pairs as equal.

Benefits

Accuracy gains over the best combined classifier of 2 to 5 percentage points on CIFAR-100 and 1 to 3 on ImageNet-1k. Published in IEEE Access and TMLR: https://doi.org/10.1109/ACCESS.2023.3298057

Technical solution

Architecture

Pairwise decomposition of the multi-class problem, with a weighted tensor-based combination layer over the outputs of heterogeneous backbone models.

Technologies

Python, PyTorch, convolutional neural networks, vision transformers, R with ggplot2 for analysis

Public Service Network Design Under Conflicting Criteria

UNIZA, University of Žilina

Planners get the full range of viable trade-offs for a regional service network in a fraction of the time exact methods need.

#Government
#Strategy - Planning

Problem

Designing a public service network means trading off criteria that pull against each other, cost against coverage against response time. Exact methods produce provably optimal answers but take long enough on region-scale problems to be impractical for planners exploring alternatives.

Solution

A modified genetic algorithm that approximates the Pareto set rather than solving for a single optimum, so planners see the range of viable trade-offs instead of one answer conditioned on weights chosen in advance.

Benefits

Computation 50% to 85% faster than the exact method across the three largest Slovak regions, with solution quality that remains practically equivalent. Published: https://doi.org/10.1016/j.trpro.2019.07.185

Technical solution

Architecture

Multi-objective optimization via a modified genetic algorithm approximating the Pareto front, benchmarked against exact mathematical programming on real regional instances.

Technologies

Java, custom multi-objective genetic algorithm with Pareto-front archiving (NSGA-II family), FICO Xpress or IBM CPLEX for the exact benchmark, Slovak regional demand and network datasets

Automated Verification of Legal Documents Against Public Registries

SingleDraft

Every party, property and legal reference in a document is checked against the public registries in one pass instead of dozens of manual lookups.

#Legal
#Legal - Compliance

Problem

Every party, property and statutory reference in a legal document has to be checked against the public registries, and lawyers were doing it by hand, one lookup at a time. The work is slow, easy to skip under deadline pressure, and a missed discrepancy surfaces at the worst possible moment.

Solution

Target information is located in the unstructured document text, matched against the commercial register, cadastre and collection of laws, and returned as a comparison showing what agrees, what conflicts and what the registry knows that the document does not.

Benefits

Manual lookup across the supported registries is eliminated, reducing verification from a sequence of manual searches to a single pass and making full checking practical on every document rather than only the high-value ones.

Technical solution

Architecture

Heuristic entity extraction over unstructured legal text, with fuzzy and multi-pattern matching against public registry APIs, results reconciled into an agreement-and-discrepancy view.

Technologies

Python, Zeep, FastAPI, regex, ahocorasick-rs, RapidFuzz, AWS S3 and DynamoDB

LLM-Assisted Legal Drafting on an Internal Corpus

SingleDraft

Drafting suggestions come from the firm's own precedent and arrive as tracked changes to accept or reject, not as text to retype.

#Legal
#Administration - Documentation

Problem

Firms hold years of precedent in their own document corpus but cannot reach it while drafting. Lawyers rewrite clauses that already exist somewhere in the archive, and general-purpose language models produce text that ignores the firm's own established wording.

Solution

A multi-branch RAG system that combines deterministic routing with LLM-driven decisions and returns its output as proposed changes to the document rather than as free text, so the lawyer reviews edits in place and accepts or rejects them individually.

Benefits

Several-fold speedup in editing and extending documents, with proposals grounded in the firm's own corpus rather than in generic language model output.

Technical solution

Architecture

Multi-branch retrieval pipeline mixing deterministic and model-driven decision points, with output structured as document change proposals for human approval.

Technologies

Python, LangChain, Pydantic, Qdrant, Langfuse, DeepL, FastAPI

Tabular Data Extraction from Mixed-Format Documents

Arles

Orders arriving as PDFs, spreadsheets or plain email text all come out as the same structured line items, without retyping.

#Retail
#Sales - Ordering

Problem

Orders and quotations arrive by email in whatever format the sender happens to use, as free text, PDF, spreadsheet, XML or HTML. Someone has to read each one and retype the line items into the system before anything can be processed.

Solution

A proof of concept extracting tabular data from any of those formats into a single structured output, using multimodal models where the content is visual and structured LLM output to guarantee the result matches the expected schema.

Benefits

Demonstrated automation of the retyping step in order and quotation intake, removing the format-by-format manual handling that had made the process impossible to automate with rule-based parsing.

Technical solution

Architecture

Format-agnostic extraction with schema-constrained LLM output, multimodal handling for visually structured source documents.

Technologies

Python, LangChain, Pydantic

Agentic Data Discovery and Visualization

Internal R&D

The same agent built three times on three frameworks, producing a like-for-like comparison rather than an opinion about which to use.

#Technology
#Data - Visualization

Problem

Answering a data question often means finding the data first, and the data is scattered across the public web in tables nobody has consolidated. The step from question to chart involves searching, scraping, cleaning and plotting, each of which is individually simple and collectively too tedious to do by hand.

Solution

A two-stage agentic system with tools for web search, web scraping, and code generation for post-processing and rendering the chart. The same system was implemented three times, on LangGraph, Pydantic AI and the OpenAI SDK, to compare the frameworks against each other on identical requirements.

Benefits

Produced a direct comparison of the leading agent frameworks under identical conditions. Pydantic AI emerged as the working compromise between the breadth of the LangChain and LangGraph ecosystem and the bare building blocks of the OpenAI SDK, with type safety as the deciding advantage.

Technical solution

Architecture

Two-stage agent, discovery followed by processing, with generated code executing the post-processing and plotting step. Built three times against different frameworks for direct comparison, with full tracing on each implementation.

Technologies

Python, Pydantic AI, LangGraph, OpenAI Python SDK, Brightdata, Pydantic, FastAPI, Logfire, LangSmith, Langfuse

Recognise your problem here?

Most of these started as a half-hour conversation about where the work was piling up.

Book a call