Ilia Bukin
Data & AI engineer, London, UK
I build the whole thing: product, API, data platform, deployment. Right now that's the data and AI engineering at a retail business, plus an AI studio I co-founded.
About
I'm the sole engineer at a retail company, where I built its data platform and the internal app on top of it from scratch. I like the messy middle of a business: working out what people actually need, then shipping the smallest thing that answers it. Alongside that I co-founded LyraGen, an AI studio, where we built Lyramed, a clinical decision-support product for a regulated industry.
Experience
Data & AI Engineer · Purebaby
Apr 2025 – Present · London, UK
End-to-end delivery ownership as the only engineer embedded in a retail business, from warehouse to the app people use every day.
- Gathered requirements from Planning, Sales, and Marketing, then designed, shipped, and now run everything in production.
- Built the company's data platform end to end: Dagster-orchestrated ingestion from 10+ source systems, 200+ dbt models with 240+ tests, and a 90+ cube semantic layer that is now the single source of truth for reporting.
- Built RetailFlow, the internal Next.js platform on top of it: real-time dashboards, project management, and team chat, with MFA, SSO, and page-level RBAC. Pentested against production, all critical findings fixed.
- Cut projected Snowflake spend by about half with per-source cost attribution and incremental models, taking the heaviest rebuild from ~380 to ~30 runs a month.
- Shipped an AI assistant and an MCP server over the semantic layer so any team can query live company data through Claude in natural language.
- Advise the leadership team on technology strategy and build-vs-buy.
Co-Founder & Full Stack Engineer · LyraGen
Feb 2026 – Present · Remote
Building and selling software for regulated industries, covering client scoping, architecture, and shipping in a two-person team.
- Co-founded the studio and shipped Lyramed, a clinical decision-support product: a Next.js app in front of a FastAPI service, secured with JWT auth, with CI/CD and tests across all three services.
- Designed the multi-agent pipeline: a triage agent routing to pharmacology, lab analysis, and imaging specialists, behind a 5-layer safety pipeline screening every response.
- Designed the GDPR data lifecycle, including DSAR deletion across the app, agent state, and object storage via the outbox pattern.
- Scoped requirements directly with non-technical clients in a regulated industry and translated domain rules into production systems.
Research Assistant · Queen Mary University of London
Sep 2022 – Nov 2024 · London, UK
Data collection and machine-learning research, on and off alongside my degrees, including a paper on hybrid retrieval.
- Built a scraping pipeline that extracted 100,000+ app-store reviews and fine-tuned a text classifier on them, improving accuracy by 15%.
- Authored a paper on hybrid retrieval, combining knowledge-graph and vector search, and benchmarked four pipelines against a shared evaluation framework.
Education
MSc Computer Science · DistinctionQueen Mary University of London
Sep 2023 – Sep 2024
BSc (Hons) Business with LawQueen Mary University of London
Sep 2019 – Jun 2022
Stack
Languages
Frontend
Backend
Data
AI & agents
Cloud & DevOps
Featured projects
View allLyramed
A 9-agent clinical AI system on the LangGraph swarm pattern, wrapped in a 5-layer safety pipeline.MoveScout
A solo-built UK property search engine aggregating OnTheMarket, Rightmove, and Zoopla with geospatial search over PostGIS.RetailFlow
Purebaby's company-wide analytics and operations platform, the app the data platform and MCP server plug into.
Latest posts
View allJev: the System One model that answers software with probabilities instead of prose
TypeSafe AI is out of stealth with Jev, a "System One model" that skips text generation and answers software with typed, calibrated probabilities at 70-500ms and $0.042/MTok input. Here's how the API works, where it fits in the agent loop, and what the evals actually show.OpenSpec: the spec layer that gets AI coding agents to plan before they code
Fission AI's OpenSpec is riding a wave: 69.5k stars, a rebuilt "actions, not phases" workflow, and a Hacker News thread that loves and hates it in equal measure. Here's what it does and what the debate gets right.Open Code Review: the deterministic review CLI that beats general-purpose agents at a ninth of the tokens
Alibaba's `ocr` splits code review between deterministic pipelines and an LLM agent, and its own benchmark says the harness around the model matters more than the model.
Contact
Open to engineering roles and to building things with people who have a problem worth solving. Email is the fastest way to reach me.