AI / Agentic Systems Architect & ML Engineer · Stavanger, Norway 🇳🇴

I design, build and ship AI systems to production — with version control, evaluation and observability, not notebook prototypes. Two years delivering production-grade AI across the full stack: agent systems, RAG pipelines, computer vision, neural-model training, AI voice agents, MCP tool servers and n8n automation. Strongest in industrial & regulated domains (computer-vision inspection, compliance) and multi-agent systems.

🌐 aiforalle.no — live portfolio: talk to the voice agents, open the demos 📍 Stavanger, Norway · 🗣️ Norwegian · English 💼 linkedin.com/in/nikita-alexeev

Evaluation-driven, observable, reproducible. I measure before I ship (precision/recall/F1, cost-per-unit), put a kill-switch on anything autonomous, and treat governance + guardrails as part of the architecture — not an afterthought. → How I architect agentic AI systems

AI-assisted development is the method, not a footnote. Claude Code, Cursor and Codex are my daily tools, from first prompt to tested code in production. I own the architecture and infrastructure decisions and the guarantee that what ships can be maintained by someone else.

Several systems below run in production and stay under closed development. Where the code isn't public, I've written up the architecture and trade-offs — and I'm glad to do a live walkthrough.

• Tilsyn.AI — AI inspection platform(own venture · runs on my own infrastructure, not publicly available) — A different approach to automated inspection in a regulated industrial domain: VLM + RAG behind an orchestrator, eval-driven quality gates and a kill-switch. The core idea is kept private, but the pattern is published as open code below — and I'm glad to do a live walkthrough of the architecture and the evaluation setup.
• vlm-rag-pipeline-demo — The readable, sanitized version of that pattern: a two-pass VLM + RAG pipeline where the model may only cite what retrieval actually returned. Grounding is enforced by the code, not by the prompt, so an unsupported claim can't survive the second pass. MIT, tested with pytest.
• NM i AI 2026 — multi-agent coordination(15th of 380, solo) — Norwegian AI Championship. Qualification: a MAPF warehouse-robot controller built as a measured optimization ladder — one-way corridors (+129%) → PIBT collision resolution (+47%) → central dispatcher (+38%) → anti-deadlock layer (+53%). Final (69 h, 3 tasks): shelf-detection CV 0.39→0.88 (ensemble + WBF + multi-scale), an API accounting agent (7 languages), and a black-box simulation. → write-up · competition repo · MAPF coordination code
• Orkestrator AI — planning cockpit(built for Stavanger kulturskole · pilot autumn 2026) · live demo ↗(access code on aiforalle.no or on request) — Teachers, ensembles, students and rooms as a drag-and-drop graph board instead of a spreadsheet, with a conflict engine that flags eight kinds of scheduling clash. This is what governance looks like in practice: the AI sits behind a pseudonymization boundary so raw names and addresses never leave the process, its auto-solver may only propose fixes the deterministic engine has already verified reduce conflicts, and every edit lands in a sandbox draft a human applies or discards. Next.js 16 · Postgres · Claude (pseudonymized) · Mapbox. → sanitized orchestration demo (typed agent-coordination backend, Next.js + tRPC + Claude SDK)
• Research-built writing skill + editorial pipeline — published in print — Deep research across several LLMs on how the best magazines actually construct an explainer, distilled into a reusable writing skill (structure, nut graf, controlling analogy, anti-hype discipline). That skill plus Claude Fable 5 produced the draft; an n8n pipeline then stripped the generative tells — a Judge agent scores the text against a living editorial policy and flags unverifiable claims, a Reviser fixes what the Judge caught, and a score gate decides whether a second pass is needed. A second workflow diffs my hand-edits against the AI original and folds the lessons back into the policy, so it converges on my voice rather than a generic one. Fact-checking stays human. The output ran as a feature explainer on MiroFish in the AI Rising issue of Sooo. Magazine (July 2026) — evaluation validated by someone else's editorial acceptance, not by my own score. → read the article ↗ · pipeline code

Conversational voice agents I built on ElevenLabs — bilingual (NO/EN), demoable in the browser:

• SIMO — in-cabin voice agent · talk to it ↗ — voice assistant for the SIMO motion-simulator platform (Heimdall Innovation).
• Stavanger Kulturskole — AI secretary · talk to it ↗ — front-desk voice agent for a municipal music school: inquiries + booking-style requests.
• Visam AI agent · talk to it ↗ — CRM-connected phone voice agent for Visam AS. It acts, not just talks: ask it to email you the price list or book a consultation (Innovation Park Stavanger or online) and it captures the lead end-to-end. Bilingual NO/EN.
• VoxBridge — live call translation(in production · novel) — real-time two-way phone-call translation: each caller hears the other in their own language, synchronously. Twilio Media Streams + OpenAI Realtime API.

• FinnFunn — AI-driven social-media automation · on Facebook ↗ — A social-automation system driven by AI agents with an n8n backend: scrapes, scores, generates and schedules content daily, with a Telegram human-in-the-loop approval step before anything publishes. Runs in production, publishing daily to its Facebook community. Multi-model (Claude/Gemini/GPT), Deepgram, Supabase, Hetzner scraper. → write-up
• BRO Norge — AI-matched work brigades · live demo ↗ — A Norwegian service company that employs Ukrainian refugees in managed micro-teams (brigades): AI matches team composition, runs onboarding and sales outreach, and handles NO/UA/EN translation across the platform. The client buys a service (one invoice, a Norwegian-speaking lead on site, no HR overhead); the worker gets a route into employment. Two-audience site (clients + workers), Next.js on Netlify.
• n8n automation at scale — 200+ production workflows across products: lead scraping & enrichment (Google Places, BRREG), cold-email outreach (Brevo), video generation (Shotstack/Pexels) and booking chatbots (Google Calendar) — with retries, throttling, idempotency and webhook routing. → patterns write-up

• claude-prompt-engineering — Context-aware prompt-generation framework: 36 techniques synthesized from Anthropic docs + research, model-specific adaptations.
• mirofish-guide ⭐ — Operator workflow, evaluation rubric & troubleshooting for a multi-agent simulation engine (GraphRAG, knowledge graph, prompt engineering).
• openclaw-roborock-cli — Agent-skill backend + CLI for scriptable robot control; room-aware, ADB fallback, CI.
• openclaw-tune-codex-plus — Cost-aware agentic ops: subscription-tier routing, quota-lifecycle tuning, security/isolation trade-off docs.
• claude-code-skill-tldraw-diagram — VS Code skill generating schema-valid tldraw architecture diagrams.
• n8n-write-human — The editorial pipeline above, as runnable code: draft → judge (a separate model scores against a living policy) → revise → deterministic cleanup of the AI tells. Judging and enforcement are deliberately split.
• davinci-resolve-mcp-v2 — MCP server for DaVinci Resolve, rebuilt modular with a confirm-gate safe-mode layer over destructive operations. Extends samuelgursky/davinci-resolve-mcp.
• screenreader-retrofit — Agent skill that retrofits an existing web app for blind screen-reader users (NVDA/VoiceOver/TalkBack): semantics, live commentary, focus management, sonification, haptics. Written while making cablegoal.com actually usable without sight.
• notebooklm-mastery-guide — Beginner-to-expert NotebookLM guide (EN/NO) plus a 27-workflow agent playbook of paste-ready prompts.
• claude-kahoot-skill — Agent skill for designing and safely importing Kahoot quizzes.

• Cable Goal — forensic 3D replay(cablegoal.com) — A browser 3D reconstruction of the disputed Norway–England World Cup goal, the shot that appeared to clip an overhead TV cable before England's equalizer. Re-kick it yourself and watch the trajectory branch on whether it hits the wire, then vote. Built solo end to end: physics, rigged 3D players, a rate-limited voting API and a live per-country dashboard. 1,800+ votes from 35+ countries, and featured in the Norwegian weekly Morgenbladet (19 July 2026). Three.js · Netlify Functions · WebAudio · Node. → Morgenbladet article ↗ · repo
• Viskr — real-time AI call coaching(novel — iOS + Android, in production) — An app that rides a live phone call and feeds you the next line on screen: real-time transcription, speaker separation, AI prompts mid-conversation. For sales calls and interviews, where the right line a second too late is the wrong line. Nobody else ships this. Approach kept private — live walkthrough on request.
• PanteMania(pantemania.no) — Norway-first mobile game with an original real-world-to-game mechanic; full-stack React Native + Express + Prisma. Collect-and-battle gameplay tied to a real national system.
• Sodarr(sodarr.com) — Coca-Cola vs Pepsi, tasted properly. A tasting app: scan the bottle or can in front of you and rate it while drinking it — five recipe axes (sweetness, bitterness, aftertaste, aroma, body), flavour tags, 1–5 verdict — and every verdict feeds a weekly Coca-Cola vs Pepsi battle scored city by city. Each tasting is also compared against your own history ("sweeter than your usual"), privately. The interesting modelling problem: three further axes (temperature, carbonation, freshness) are kept brand-independent so the same event yields a second signal — how well the bottle was stored — without preference contamination (Norway is an extreme Pepsi Max market; raw scores would drown it). Geo queries and score aggregation run inside Postgres (RPC + scheduled jobs), so it runs on a free tier. React 19 + TypeScript + Supabase + MapLibre GL, 11 languages; built solo.
• VibeVox(vibevox.io) — Windows voice dictation/translation; on-device ASR — choice of NVIDIA Parakeet (NeMo, INT8) or OpenAI Whisper — packaged installer.
• SIMO(Heimdall Innovation) — Unity 6 motion-simulator platform: DJI O4 video → washout-filter motion cueing → motorized chair, with an in-cabin voice agent (above). Shipping as SIMO — Simulation Redefined ↗. → video of the rig in motion ↗
• PrisSnap(prissnap.no) — AI pricing for FINN.no (web + Android): Gemini Vision → price estimate + ready-to-post Norwegian listing in seconds.
• Byggekostnad Pro · live demo ↗ — Norwegian construction-cost estimator: a guided step-by-step flow (project, dimensions, ground, structure, technical, quality) updates the budget live as you pick timber vs concrete, insulation class, finish level — with saved calculations, company profiles, exportable PDF cost reports and an inline AI advisor for building-cost questions. React + TypeScript + Supabase.
• SkiPuls.AI(skipuls.no) — Privacy-first crowd estimation for ski resorts: lightweight CPU object-detection at the edge counts people on a slope or in a lift queue, each frame masked to a region of interest and discarded on the spot — no facial recognition, no stored images, only the headcount feeding a live dashboard. FastAPI + Postgres + Next.js + Docker.
• JobSwap — commute-matching ML platform(Visam AS · R&D approved under SkatteFUNN / Research Council of Norway) · myjobswap.com · my v2 build ↗ — An AI platform that cuts commutes by matching "mirror twins" and chaining job/home swaps (Exchange Jobs, Shorten Commutes, Save the Planet — launching on Kickstarter). I built the ML system: 21 ONNX-exported models — a spatio-temporal traffic GNN, a Siamese compatibility network, an XGBoost LambdaMART ranker and an ILP chain solver (kidney-exchange algorithm) — trained on 1M synthetic profiles (privacy by design), plus a v2 React + Mapbox demo app over a Python backend. → ML write-up · growth pipeline

• GlobalPrice · costglobe.com ↗ — Electricity/fuel/transport price tracker across 36 countries on an interactive WebGL globe with markerless head-tracked perspective — webcam face-tracking shifts the 3D viewpoint in real time (head-coupled "fish-tank" parallax, no markers). Node/Express backend with resilient per-source scrapers + a validated API.

What I build with — using AI-assisted development as the working method (see above), so read this as the stack I ship and maintain, not as hand-written-from-scratch fluency.

Languages & runtimes Python · TypeScript/JS · SQL · Bash · Kotlin · C#/Unity · Go · R AI-assisted dev Claude Code · Cursor · Codex · agent skills · MCP AI/ML Claude · OpenAI · Gemini/Gemma · VLM/multimodal · RAG (pgvector, Pinecone) · NN training/fine-tuning · GNNs · learning-to-rank (LambdaMART) · object detection · STT (Parakeet, Whisper, Deepgram) · OpenAI Realtime · quantization (FP8/INT8) · ONNX · vLLM · eval (P/R/F1) Agents MCP servers & tool-use · multi-agent orchestration · voice agents (ElevenLabs) · n8n orchestration · human-in-the-loop · ReAct / plan-execute · guardrails · multi-provider LLM gateway Backend/data FastAPI · Next.js · Node/Express · REST/WebSocket · Twilio · Brevo · Google APIs · Supabase/Postgres (pgvector, RLS, Alembic) · n8n Infra Docker/Compose · GPU orchestration (vast.ai) · Vercel/Netlify/Hetzner · GitHub Actions

Project Lead & Developer @ Visam AS, Innovation Park Stavanger — industrial, maritime & energy clients.Certified: Anthropic — Claude Code in Action (2026).

_{Made with ♥ by Alexeev Digital Lab}