Sergei Audzeichyk
My work begins where the impressive AI demo ends.
I work as an AI/ML engineer and AI product developer. My work sits at the intersection of machine learning, software engineering, law, education, and applied research.
Also published as Siarhei Audzeichyk, Sergey Avdeychik and Сергей Авдейчик.
How I came to this work
One project began with a conversation with a human-rights lawyer. He described hundreds of pages of material, procedural deadlines, formal requirements, and a weak but critical signal that could not be missed.
As an engineer, I recognised the structure of the problem. In computer vision, a system learns to find a meaningful pattern in a noisy image. In legal casework, the ‘image’ becomes a body of documents, facts, dates, positions, and sources. The cost of error is no longer defined by a model metric, but by the consequences for a specific case.
That encounter led to a sustained interest in AI systems for complex professional work. The central question is not whether a model can write persuasive text—it can. What matters is whether the system can show what the text rests on, where uncertainty remains, and when it must stop before a professional decision.
What I do
I design AI copilots, RAG systems, agent workspaces, research pipelines, and intelligent-automation tools—from version-aware retrieval over legal corpora with citation controls to open research in which data and code are published alongside the conclusion.
How I work
I begin with the user's task, not the model name. First I establish which decision the user makes, which sources are admissible, what an unacceptable error looks like, and what can be measured before deployment. I then build the smallest live prototype around one workflow.
Tests, logging, and reproducibility arrive with the first working version, not at the end. If the hypothesis does not survive scrutiny, it should be changed or closed—not concealed beneath another layer of agent orchestration.
Why I publish openly
Open code and data are not a way to display activity; they are a way to expose the method. Open work does not prove that there are no errors. It makes errors visible and allows another person to improve the result.
Technology approach
I work with Python, FastAPI and Quart, MongoDB, Redis, React and Next.js, PyTorch and TensorFlow, and models from OpenAI, Anthropic, and Mistral. I prefer official SDKs and transparent custom pipelines to heavy frameworks when abstraction hides data, state, or the cause of failure.
The stack is a means. Architecture is determined by the problem, access rights, cost of change, and the team's ability to maintain the system after the pilot.
Persuasive text is the cheapest thing the age of generative AI has produced. Value begins when we can show how a conclusion was formed, which checks it survived, and why a person decided to rely on it.