Start with the question
I don't reach for a model before I know which decision it has to support. Get the question right and the system usually designs itself.
Available for new work
Head of AI.
I'm Nikolay Baakh. I build production AI systems and lead the teams that ship them: agents, RAG knowledge bases, and the platforms around them. Before AI, I built and ran data departments as Head of Data across iGaming, B2B SaaS, and affiliate marketing.
- 3+Years in AI
- 20People led
- 3Teams built from zero
Now
Updated September 2026Head of AI, on contract for a client in Cyprus. On their analytics platform I'm building a natural-language-to-SQL agent over Postgres and ClickHouse, a pgvector RAG knowledge base, and the React admin UI around it.
On the side: writing the next data research video and refining Epstein AI on top of the full court document archive.
01/Work
Things I built, and keep running.
Built in the open, all on GitHub.
Open source
Epicurus
A self-hosted, modular, local-first personal-assistant platform — an AI agent plus pluggable sidecar modules for calendar, notes, tasks, mail, chat, knowledge base, and storage. Runs on your own machine under Docker. Early stage, open source under AGPL-3.0.
github.com/baakhoff/epicurusBuilt in public
Homelab
Production infrastructure, learned on spare hardware and documented as it's built. Three mini PCs run a k3s cluster with replicated Ceph storage, reconciled by Flux, behind passkey sign-in and watched by Prometheus, Grafana and Loki.
github.com/baakhoff/homelabOpen source
BiD
The missing Linux mixer for Audient iD audio interfaces: mixes and cues, routing, meters and the monitor section, while audio keeps playing. A fork of MixiD, reworked to talk over the device's spare DFU interface so the kernel keeps the sound card.
github.com/baakhoff/BiD02/Research
Deep dives, published as videos.
Real datasets, real methodology, and a clear answer at the end of each one. Every video has a written breakdown.
EP. 01 — RAG / LLM
Hear me out... Epstein AI It cites the page.
A RAG agent over the public archive of Epstein court documents. Ingest, OCR, chunk, embed, retrieve. Ask anything; it answers with citations to the source file.
Read
EP. 02 — Python / Statistics
I analyzed 11,463 D&D Monsters, all Spells and Items. You are playing wrong.
A statistical deep-dive into every creature, spell, and magic item ever published for Dungeons & Dragons, and the balance most players never question.
Read
EP. 03 — SQL / Web scraping
I Downloaded All Data from Rule34.XXX (For Science)
Data engineering and statistical analysis on one of the internet's largest image-board datasets. Same rigour you'd apply to any enterprise source, on a much weirder one.
Read
EP. 04 — scikit-learn / Space
To Predict the Apocalypse I Stole NASA's Asteroid Data
A predictive model on NASA's public Near-Earth Object dataset. Orbital mechanics, Python, and one question: how worried should we actually be?
Read
03/About
From data to AI.
My path wasn't a straight line. I started out running social media campaigns, got obsessed with the numbers behind them, taught myself SQL and Python, and talked my way into a data analyst role at a retail company.
Before AI, that became a data career: from analyst to Head of Data, building departments from zero in iGaming and affiliate marketing, with ClickHouse warehouses, real-time pipelines, and teams of up to twenty. AI crept into every layer of that work, and now it is the work: as Head of AI I build production AI systems end to end, most recently a platform with agents, a RAG knowledge base, and the security hardening around it.
- Belgrade, Serbia
- English · Russian
- Remote-first, CET
Build for trust
A model no one trusts is worthless. Explainability, documentation, and stakeholder buy-in get real time. Adoption is half the ROI.
Iterate fast, document well
Good AI work ships in weeks, not quarters, and leaves behind systems your team can run after I'm gone.
Experience
-
May 2026 — Present
Head of AI · Analytics platform
Confidential client, Cyprus · remote contract
-
Sep 2025 — Apr 2026
Head of Data Department
B2B iGaming tech provider, Lisbon
-
Jan 2025 — Sep 2025
Head of Data Department
iGaming affiliate company, Netherlands
-
Aug 2023 — Jan 2025
Team Lead → Product CAO
01tech, Belgrade
-
Oct 2022 — Aug 2023
Data Analyst → Lead Data Analyst
ПРО ДИЗАЙН, St Petersburg · Ocean Digital, Almaty
-
Jan 2020 — Jan 2022
Marketing Specialist
Self-employed, Russia
Stack
- LLM agents
- RAG
- Python
- pgvector
- MCP
- Ollama
- FastAPI
- React
- TypeScript
- ClickHouse
- PostgreSQL
- SQL
- Docker
- Kubernetes
- GitLab CI
- Keycloak
- Grafana
- Prometheus
- dbt
- Airflow
- pandas
- scikit-learn
- REST APIs
- Swagger
04/Services
End-to-end AI work, fitted to what you're actually doing.
AI strategy, production systems, and hands-on leadership for a critical quarter. You get the depth of a Head of AI without committing to a full-time hire.
AI strategy & roadmap
A senior outside view of where AI actually pays off in your business: which use cases to ship first, what they need from your data and team, and a roadmap leadership can fund.
Use-case selection · Build vs buy · Roadmaps
Production AI systems
RAG knowledge bases, agents, and LLM integrations, from a scoped proof-of-concept to production with SSO, access control, observability, and monitoring.
RAG · Agents · LLM APIs · MCP
Ask-your-data agents
Agents that answer questions straight from your warehouse in plain language, with schema awareness and safety nets on every query. Built over Postgres and ClickHouse.
Text-to-SQL · ClickHouse · PostgreSQL
Security & reliability
Hardening AI products before real users find the gaps: access control, SSRF and IDOR, auth flows, concurrency bugs, and the tests that keep them fixed.
SSO & RBAC · Security fixes · Observability
Data foundations for AI
The groundwork AI depends on: warehouse design, pipelines, and data quality, so models and agents work from clean, well-governed data. This is where my career started.
dbt · Airflow · ClickHouse · Data quality
AI adoption & team enablement
Frameworks, workshops, and hiring support that get a team using AI every day. Past frameworks cut analysts' task time and roughly halved a data team's pipeline shipping time.
AI usage frameworks · Workshops · Hiring
How we can work together
2 – 4 weeks
Focused sprint
A time-boxed engagement for one well-defined problem: an audit, an agent, or a proof-of-concept that needs to ship fast.
Most popular
3+ months
Monthly retainer
A set number of days a month across strategy, architecture, or implementation. Effectively a fractional Head of AI.
1 – 6 months
End-to-end project
Fixed scope with defined deliverables and milestones. Best for production AI systems and the infrastructure under them.
Open to hire
Full-time role
Head of AI, in-house. Remote-first or hybrid in EU time zones; iGaming, B2B SaaS, or affiliate.
05/Contact
Let's put AI to work.
A specific project, a general question, or just sounding out whether we'd be a fit. I read every message myself.
- Reply within one business dayEvery message gets a personal answer.
- Free 30-minute scoping callA no-pressure conversation before any proposal.
- NDA on requestHappy to sign before anything sensitive is discussed.
- Quick startFocused sprints can usually begin within one to two weeks.
Taking on consulting engagements alongside my current contract, and open to full-time Head of AI roles. Remote-first from Belgrade, and happy to travel within Europe for workshops.