Русская версия

AI Product Engineer

I turn ambiguous problems into working AI products

I work out what’s actually needed, define a sensible first version, build and test it with AI coding agents, ship it, and iterate on real feedback.

Get in touchSee selected work
  • 3 live AI demos
  • 2 production Telegram products
  • 1 public beta
  • End-to-end product ownership

Selected work

Three products, three kinds of proof: AI analysis of business data, AI answers you can verify, and a real business workflow.

RISKAI analysis

Company due diligence with an explained risk signal

Problem
Before a deal, a business needs to understand its counterparty, identify material risks and document the basis for a commercial decision.
What I built
Enter a Russian company tax ID: the service collects public registry and business data, applies an expert methodology to calculate a traffic-light signal, and uses AI to explain the facts, gaps and follow-up checks.
My role
I built the product: workflow, architecture, data integration, backend, AI pipeline and frontend. A client-side domain expert supplied the methodology; Legicon later reused the core.
Status
Live public demo for Russian companies. The interface is in Russian; a saved example report opens immediately.
Read the case studyOpen live demo

CORTEXVerifiable AI answers

Answers from long videos you can verify against the original recording

Problem
Facts in a two-hour interview are hard to find and harder to trust: a summary alone cannot be checked.
What I built
The system finds relevant passages by meaning and builds an answer grounded in them. Open a supporting source fragment in one click to check the answer against the original recording, without scrubbing through the video.
My role
I split the transcript into meaningful passages, represented them as embedding vectors and built a semantic vector index for RAG retrieval. It finds passages by meaning rather than keyword matches and links the answer to the exact moment in the original recording.
Status
Live demo on a real two-hour founder interview.

The interface and the interview are in Russian; the evidence links work the same in any language.

Read the case studyOpen live demo

STOKCENClient project · Business workflow

From an appliance enquiry to a tailored offer

Problem
Customers need to choose several appliances, each with different specifications. The retailer needs those requirements in one structured request.
What I built
The bot and Mini App gather category-specific requirements for a chosen store. One manager prepares an offer, with customer choices and follow-up kept in the request context.
My role
I designed and built the product end to end: structured intake, routing, Mini App staff tools, offers, conversation, testing and deployment.
Status
In production for the client.

Russian-language interface. This is the client’s live sales channel; please do not submit test requests.

Read the case studyOpen the bot

How I work

  1. ProblemStart from the business or user problem, not from a technology.
  2. UnderstandingFind out what is really needed: users, data, constraints and what “done” means.
  3. ScopeCut it to the smallest version that proves value, and say what is left out.
  4. Working productDesign just enough architecture and build a product people can actually use.
  5. ValidationTest it, ship it to a real environment and change it based on what real use shows.

I build with AI coding agents, mainly Claude Code and Codex. They speed up implementation. I own the product framing, the architecture, the constraints, code review, testing, deployment and the result.

More shipped products

  • MedRecall

    Live synthetic demo; separate closed pilot deployed

    Helps a clinician reconstruct the history across calls: see what changed, find outstanding follow-ups and check an answer against the transcript and recording.

    Read the case studyOpen live demo

  • Acoustic

    Live paid product · Trial credits available

    A long recording becomes a speaker-labelled transcript, named voices and notes on demand with links to supporting audio. Users can verify findings and export a named transcript for further analysis.

    Read the case studyOpen the bot

  • Dating Assistant

    Public beta

    Designed to reduce the effort of starting and keeping track of dating conversations: profile-based openers, one-off screenshot replies and replies from saved dialogue history. Three workflows, each with its own context.

    Read the case studyOpen the bot

Engineering tools / Evaluation labs

  • STT Lab · Darwin

    A workspace for reviewing transcripts and speaker attribution against the original audio. It supports side-by-side comparison, human judgments on disputed segments, and repeatable experiments across speech pipelines.

  • Prompt Lab · Dating Assistant

    An internal evaluation workspace for comparing models and prompt versions on the same test cases. Saved runs, cost and latency data, and human review help catch regressions before release.

Ways to work together

I can join a team, work on a long-term contract, or build a scoped product for a business.

  • Team role

    A full-time role in a product team, owning AI features or new products from problem to production.

  • Long-term contract

    An ongoing engagement with a startup or product company, working inside your process and tools.

  • B2B project

    A scoped build for your business: from an unclear request to a working, deployed product.

About

I am a hands-on engineer who takes product ownership. I work across frontend, backend, data, LLMs and deployment, and I am comfortable owning a product alone or in a small team.

Most of my products so far were built for Russian-speaking users and businesses, so their interfaces are in Russian. The work itself — turning an unclear need into a reliable product — does not depend on the market.

Contact

Tell me about the problem, the role or the project. Email or Telegram, whichever is easier.

← Swipe →