A historical reading needs the source, territorial definition, date, and transformation description together.
Legal AI · agent systems · applied research
I build AI systems that do not have to be taken on trust.
For law, research, and complex professional work—where a persuasive answer is not enough. Every material conclusion should lead back to its source, data version, calculation, and human decision.
Sergei Audzeichyk is an AI/ML engineer and AI product developer.
How I work →POSITION / 01Models are useful, but their value is determined by the architecture of work around them.
The cost of error
Persuasive text has become cheap. A verifiable result has not.
A large language model can produce an opinion, review, or architecture recommendation in seconds. In professional work, however, the most expensive question begins after the answer.
Which document was the source? Which version applied at the relevant time? What did the system fail to retrieve? Where do the data end and interpretation begin? Which error cannot be missed? Who makes the final decision?
I design not generators of polished answers, but working environments in which the path from source material to conclusion, verification, and human responsibility remains visible.
The model may be strong. The process must be stronger than the model.BY Maps / open research
Demographic change is not explained by one figure, but by connected views.
The project's original visualisation keeps observations, estimates, and conditional scenarios distinct.
Change outside major cities must be read with the series’ boundaries, definitions, and comparability in view.
Urban concentration is a research question that requires an explicit comparison of places and periods.
Scenarios are conditional models with assumptions, not a description of a future state.
Different fields—one engineering question
How do we turn a plausible model output into work that can be checked and continued?
I work across law, data, research, and education. The domains differ, but the problem recurs: a model should strengthen professional judgement rather than replace it, while leaving an intelligible record of the work.
Open research · Research framing, ETL, modelling, visualisation
BY Maps
What happens to a country when its population has been concentrating around a single point for decades?
A public research material on demographic change, combining a map, research questions, and conditional scenarios. Results should be read with their source, transformation, and limitation.
What can be verified: The research site and repository lead to material and methodology; this card does not repeat figures without a versioned dataset reference.
External repository reference · This site does not claim a role or current project status
POISK-IGR
How can external material be read when repetition in secondary sources is not independent confirmation?
An external repository linked for independent inspection. This site does not make claims about its data, verification, contributors, or current status.
- entity
- claim
- source
- verification
What can be verified: An external repository and entry point are linked for inspection; its details require independent verification in that repository.
Research pilot · Data engineering, agent systems, reproducibility
NormaLab
How can we measure not what a legal rule says, but how it is actually applied?
A research pipeline connects a corpus of court decisions with classification, citations, counterexamples, and legal review.
- question
- corpus
- citation
- conclusion
What can be verified: The methodology, event model for interpretive activity, and verification history are public.
Architectural concept · Architecture, exchange protocol, oversight model
Legal Copilot Ukraine
What if a national Legal AI system were a network of personal workspaces rather than a single platform?
An architectural concept for exchanging verifiable work packages: sources, versions of legal rules, checks, and rejected hypotheses.
- source
- version
- check
- work package
What can be verified: A public architecture and the related open minius prototype for a personal legal workspace.
Open source · Workflow architecture, implementation, verification
TwinLoop
How do you stop an agent from declaring its own work complete?
Implementation and verification are separated: the specification is the contract, an independent loop looks for evidence of completion, and CI identifies the exact version that passed.
- specification
- implementation
- independent check
- CI
What can be verified: An open repository containing the workflow, tests, and verification history.
How I work
I do not start with the model.
First, we need to understand the decision a person makes and what happens if the system is wrong. Only then should we choose a model, retrieval method, agents, or interface.
- 01
Which decision needs to improve?
Not ‘adopt AI’, but shorten a specific delay, improve retrieval coverage, reduce a defined risk, or give a professional a new way to analyse the work.
- 02
Which sources are admissible?
I establish provenance, version, date, access rights, and corpus boundaries. If a source cannot be named, the system should not make a strong claim.
- 03
What can be delegated to the model?
The model receives a bounded role: extraction, classification, relationship discovery, or preparation of a hypothesis or draft. Decisions with legal, financial, or reputational consequences remain human.
- 04
How should the system fail under test?
Tests include not only convenient examples, but plausible wrong documents, superseded versions, incomplete data, edge cases, and competing explanations.
- 05
What will the next professional receive?
Not one final file, but an intelligible work package: sources, calculations, decisions, limitations, and instructions for continuing or checking the work.
The aim is not an infallible model. It is a process that detects an error before the error becomes a decision.
Writing and research notes
I write about what happens when the demo ends.
I am interested not in another model release, but in what follows: how AI meets law, data history, institutional rules, responsibility, and real engineering work.
When Intelligence Got Cheap: What Is Expensive Now
An essay on cheap applied intelligence, the reorganisation of work, and the value of intent, judgment, relationships, and new tasks.
Read the essay →AI Has Hacked the Code of Human Civilization
An original analysis by Sergei Audzeichyk of questions raised by Yuval Noah Harari’s lecture about agency, institutions, language, and human judgment.
Read the original analysis →Pressing Record Is Easy. Explaining It to a Regulator Is Harder
Google Meet, Zoom, Fireflies.ai, Otter.ai, and Gong can place a single conversation on a path through several systems—turning it into a transcript, an AI summary, a CRM record, and part of an organisation’s memory. In 2025, the French data protection authority fined a call-centre company €250,000 for infringements involving data minimisation, retention, and security. An interactive publication about seeing the consequences before recording begins.
Read the publication →A country concentrating around its capital
A long time series can reveal structural change, but only with explicit definitions, sources, and a clear boundary between history and scenario.
A swarm instead of a platform: a new architecture for national legal systems
An architectural thesis: legal infrastructure can use many controlled agent workspaces when they exchange verifiable work packages through a shared contract.
Research
I publish results so they can be challenged on their merits.
Good research does not end with an attractive chart. Readers need sources, definitions, code, assumptions, and a clear boundary between observation and interpretation.
Open research and data →For teams and organisations
You do not need another AI tool. You need to know which work it can be trusted with.
Work can begin with a process diagnostic, architecture design, or bounded pilot. The team receives intelligible material, quality criteria, and a decision on the next step.
Explore ways of working together →Profile
An engineer, researcher, and product builder working across disciplines
I work at the intersection of machine learning, software engineering, law, and applied research. I build AI copilots, RAG systems, agent workspaces, analytical pipelines, and educational tools.
I am drawn to problems in which engineering speed alone is insufficient: the domain, the cost of error, and the way another person will verify the result all matter.
More about me →Describe where the work currently loses time, quality, or evidential integrity.
For a first conversation, a short account of the workflow, sources, cost of error, and intended outcome is enough. I will suggest the shortest route to testing the hypothesis—or say plainly if AI is not the right tool.