RAG / Agent Proof-of-Value
For companies that want proof of value before investing in a platform.
What you receive
- A working demonstrator on your data
- A response-quality evaluation dashboard
- A go/no-go decision report
AI Engineering / Data ArchitectureWork from question to delivery
Datarmination
I design and build AI systems: agents and RAG, evaluation, data architecture and knowledge graphs. I help test ideas, improve prototypes and take them into use.
sources / data / contextconcepts / relations / rulesinterface / test / release
A workflow diagram, not an automated process.
01
Research
Sources, synthesis, verification
02
Data
Models, relationships, quality
03
AI
Prototypes, evaluation, oversight
04
Product
Interfaces, testing, delivery
01Services
A defined scope. A tangible result.
From testing an idea to production delivery and ongoing work. Six ways I can join your project.
LLM Observability Sprint
For teams running LLMs in production with no visibility into quality and cost.
What you receive
- Observability (Langfuse / OpenTelemetry)
- LLM-as-a-Judge evaluation with an audit trail
- Cost per request
AI Architecture Blueprint
For boards and CTOs ahead of an investment decision.
What you receive
- A target-architecture document
- An implementation backlog
- A TCO estimate
PoC-to-Production
For DS teams with notebooks that need to deliver repeatably.
What you receive
- Reproducible pipelines (Dagster / SQLMesh)
- Data validation and CI/CD for ML
- Full auditability
Entity Resolution & Knowledge Graph Sprint
For companies whose supplier, product, customer and technology dictionaries have drifted apart across systems.
What you receive
- Master-dictionary deduplication and canonicalization — from deterministic to semantic matching
- A controlled vocabulary: decision-grade definitions, scope notes, a relation typology
- A graph schema (property graph or SKOS/RDF) with structural validation in CI
AI Engineer on your team
For companies that need a senior GenAI engineer on an ongoing, flexible basis.
What you receive
- Ongoing access to a senior AI engineer
- Architecture and code reviews
- Team mentoring
Timeframes are indicative. We agree on scope, schedule and pricing after discussing your data, integrations and acceptance criteria.
02Technology stack
What I build with.
I choose tools to fit the data, task and team infrastructure. This is the toolkit, organised by application.
Agentic systems
Task flows, tools, response validation and human checkpoints.
- LangGraph
- OpenAI Agents SDK
- Pydantic
- Structured outputs
RAG and source research
Document search, combined retrieval methods and checking the evidence behind answers.
- Hybrid retrieval
- BM25
- Reranking
- Deep-research loops
Taxonomies and knowledge graphs
Controlled vocabularies, relationships between concepts and resolving inconsistent names.
- SKOS
- RDF / SPARQL
- NetworkX
- SFIA / ESCO / O*NET
Evaluation and observability
Execution traces, test datasets, response evaluation and cost per request.
- Langfuse
- OpenTelemetry
- LLM-as-a-Judge
- Multi-model consensus
Models and integration
Matching models to tasks, comparing quality and controlling inference costs.
- LiteLLM
- Claude / OpenAI / Gemini
- scikit-learn
- XGBoost
Data and search
Data schemas, processing, indexes and interfaces to existing systems.
- PostgreSQL / pgvector
- Qdrant
- Python / SQL
- FastAPI
Engineering foundations
Code, environments, tests and reproducible releases.
Python / SQL / pandas / NumPy / Git / Docker / CI/CD / uv / ruff / pytest
03Selected projects
Examples from the practice.
All projects shown here are self-commissioned. Each applies the same craft to a different subject: sources, structure, interfaces and quality checks.
01
Public project · ongoing operation
Skills Intelligence
A skills atlas, AI glossary and role profiles in one public workspace. Readers can explore knowledge from different angles, with sources and assumptions close to the conclusions.
What I built
Explore Skills Intelligence (new tab)The service operates on an ongoing basis: the atlas, glossary and role profiles are connected to a regular editorial workflow covering new signals in AI, skills and changes in work. Material receives editorial review, and validators check data consistency before publication.
Scope and limitations
- Starting point
- Knowledge about skills and changes in work is scattered across reports, dictionaries and models. Relationships and differences in meaning are hard to see.
- Key decision
- I connected concepts through stable identifiers and explicit relationships. Sources, methodology and quality checks are part of the product.
- My role
- Knowledge model, data processing, interface, editorial work and release process.
- Context and limits
- An independent research product. Maps and assessments should be read alongside their methodology and source coverage.
02
Public project
Mazury Rowerem
A guide to the Masurian Cycling Loop. It connects a map with route sections, places and trip plans. Core materials also work offline.
What I built
Explore the cycling guide (new tab)An interactive guide with day-by-day plans, GPX files and printable material. Research turned into a tool to take on a ride.
Scope and limitations
- Starting point
- Planning a trip means piecing together maps, route descriptions, transport and places to stop. On the road, answers also need to be available offline.
- Key decision
- I linked route segments to a map and elevation profile. Uncertain information has a visible status, and the guide supports offline use.
- My role
- Route data model, research, maps, interface and offline features.
- Context and limits
- An independent guide. Conditions and services can change; information awaiting confirmation is marked.
03
Public project
Nazi Report
A multilingual archive documenting perpetrators of Nazi crimes and structures of responsibility. Its educational purpose calls for particular care with sources, language and context.
What I built
Visit the archive (new tab)A public archive of profiles, relationships and events. Readers can move from an account to its sources and methodology.
Scope and limitations
- Starting point
- Biographies, relationships and events must retain their sources and meaning across languages. An editorial error can alter the meaning of a historical account.
- Key decision
- I built a shared data structure and consistency checks across language editions. Validators block publication when they detect missing requirements.
- My role
- Archive structure, research, interface and publication controls.
- Context and limits
- An educational project based on documented sources. Available material and established findings define the scope of each account.
04
Local tool
HELM
A local tool connecting projects, tasks, decisions and memory. It helps recover the context of your work and find the next step without searching through another pile of notes.
What I built
Discuss a similar toolChanges have a history, information retains its sources and proposals need approval. Backups, export and restore are part of the application.
Scope and limitations
- Starting point
- Projects and decisions were scattered across files and exports. An old note could easily be mistaken for a current decision.
- Key decision
- I separated sources, observations and approved changes. The application can suggest a next step, but does not take it on the user’s behalf.
- My role
- Data model, browser application, local server, imports, change history and tests.
- Context and limits
- An independent local tool, not a public service. The illustration shows the workflow, not a private database or an interface screenshot. Hiding sensitive entries on screen is not encryption.
05
Public project · ongoing operation
Wspólny Punkt
A review of conservative press across countries. Source selection, translation and synthesis produce Polish-language reports that show shared events and differences in interpretation.
What I built
Explore Wspólny Punkt (new tab)The service operates on an ongoing basis: successive batches of material are collected, translated, connected into story threads and editorially reviewed. Readers can follow each report back to its sources.
Scope and limitations
- Starting point
- The same events are reported in different languages and contexts. Comparing perspectives requires preserving sources and distinguishing facts from interpretation.
- Key decision
- I connected source material into story threads and separated public reports from the editorial workspace. Each synthesis retains its source basis and the scope of material reviewed.
- My role
- Product concept, source selection and modelling, editorial workflow, reader interface and editorial workspace.
- Context and limits
- A review of a selected strand of the press, not the whole debate. Comparing perspectives and showing reading coverage helps assess a report, but does not replace evaluating the sources themselves.
06
Public project
Atlas Małych Odkryć
150 ideas for small websites about Poland and an Atlas of 344 Kondracki mesoregions. Beyond browsing a catalogue, visitors can open the microsites and compare how the same subject works across versions V1–V4.
What I built
Explore the Atlas of Small Discoveries (new tab)A catalogue connected to a map, sources and evaluation criteria. Versions V1–V3 explore different ways to present the same subject, six topics have been expanded into full V4 editions, and /kondracki/ provides 344 region studies and 688 GIS exports. A separate workshop opens up the code of V2.
Scope and limitations
- Starting point
- A list of ideas does not show whether a subject can become a useful website. Content, form and interaction need to be tested together.
- Key decision
- I connected research with working microsites. Preserved versions make it possible to compare concrete design choices rather than judge an idea from its description alone.
- My role
- Concept, research, catalogue criteria, direction of V1–V4 microsite generation, preparation of 344 Kondracki region cards, editing, interface, map, and comparison of versions and code.
- Context and limits
- Microsites V1–V3 are concepts and demonstrators, not a set of verified guides. Sample data needs confirmation before use in the field. Code analysis covers V2.
04My approach
How the work happens.
Every project takes its own path. These principles help me keep the work on course.
- 01
Start with a decision
We establish who will use the result, what gets in their way today and how we will recognise an improvement.
- 02
Examine the foundations
I inspect the data and sources, separate facts from assumptions and compare the idea with a simple solution without AI.
- 03
Build it and test it
A working part takes shape. We test mistakes, missing data and the points where a person needs to decide.
- 04
Leave an understandable system
Code, documentation and setup instructions are part of the result. What we learn determines the next step.
05Behind the practice
Piotr
Pytlarz.
Source research. Data engineering. Product development.
Visit Piotr’s Cabinet (new tab)I am interested in the space between disciplines. In one project I organise concepts; in another I connect maps with route data; in a third I work out how to publish responsibly. I combine engineering with the patience of a researcher and an editor.
Datarmination is my independent practice. I take work from the first question to a working interface. I like to understand what something will be used for, and to build it so the next person can understand how it works.
Datarmination · data, AI and product designBefore you get in touch
01When is a good time to reach out?
As soon as you can describe the problem, even if the solution is still unclear. I can also help assess an existing prototype or define the next phase of work.
02Does every project need AI?
No. I first look at what simpler methods can achieve. A model makes sense when it adds testable value to a specific task.
03How will we agree on scope and cost?
We start with the problem, available material and desired result. I then propose a first phase and a way to assess what it delivers.
04What will I receive at the end?
The agreed result with the relevant code, documentation and setup instructions. We agree on rights and ongoing maintenance before work begins.
06Contact
Tell me where you are starting.
Tell me what you want to achieve, what is getting in the way and what material you already have. A few sentences are enough to find a starting point.
contact@datarmination.com
No form. Your message comes directly to me.