Practice
AI Integration Engineer
LLM pipelines that run unattended and are still right the next morning
I build production systems around language models — the unglamorous part: orchestration, validation, retries, cost, and the evaluation that tells you whether the thing still works. An automated content pipeline that generates, checks, edits and translates with no human in the loop. A streaming transcription module that turns speech into a structured document. Full stack underneath, so the model stays a component instead of becoming the architecture.
What I can do
- Design an LLM pipeline with the boring parts included: validation, retries, idempotency, and a cost ceiling.
- Get structured, schema-checked output out of a model instead of prose you then have to parse.
- Build retrieval over your own documents, and know when it hands back the wrong passage.
- Wire speech and documents into a system: transcription, extraction, and a generated artefact at the end.
- Set up evaluation, so that changing a prompt is a measurement and not a feeling.
- Say plainly where a model does not belong, and write ordinary code there instead.
Where this came from
- 2025 —
ASMB Tech
Litseller: an automated LLM orchestration pipeline that generates, validates, edits and translates articles into eleven languages and publishes them with no human in the loop. Auto Inspect: streaming voice transcription structured into a generated PDF report.
- 2023 — 2025
Yandex
High-load B2B modules and the integration layers between microservices and external analytics.
- 2019 — 2022
Sber
Parsing and structuring unstructured data inside enforcement documents, cutting manual error.
Stack
- AI and language models
- OpenAI API ·Claude API ·LLM orchestration ·Prompt engineering ·Structured output ·RAG ·Speech transcription ·Evaluation harnesses
- Backend and data
- Node.js ·TypeScript ·PostgreSQL ·Supabase ·Queues ·REST API
- Frontend
- React ·Astro ·Next.js ·Streaming UI
- Infrastructure
- Docker ·GitHub Actions ·Vercel ·Playwright
About this on the site
Working together
Write to me if this is the shape of the problem you have.
Get in touch