MWM-AI CONSULTING · 2026
A short note

Quiet, careful engineering for people building with agents, MCP servers, and RAG.

I take on a small number of consulting engagements — architecture reviews, retrieval pipelines, MCP server design, and the occasional trading system. My work tends to favour the simplest thing that ships, proved with tests rather than promises.

On offer

Three kinds of work.

I.

MCP server architecture

Design, review, or build production MCP servers with FastMCP. Useful when your team has a prototype that works for one user but falls over on the second, or when you're deciding between an MCP server and a plain HTTP API and want a second opinion.

  • Tool design and error-handling patterns
  • Multi-tenant and multi-server orchestration
  • Integration with Claude Desktop and Claude Code
  • Migration from ad-hoc tool endpoints to MCP
II.

Agents & RAG pipelines

Retrieval systems that actually retrieve. Vector databases (Qdrant), graph knowledge (Neo4j), and local or hybrid LLM stacks. I'm particularly useful when the problem is "search quality is poor and nobody knows why" or "we need this to run on-prem."

  • Chunking, embedding, and reranking strategy
  • Hybrid search (semantic + keyword + graph)
  • Multi-agent coordination and handoffs
  • Local-first deployments for privacy or cost reasons
III.

Trading system engineering

Pine Script v6 strategies, realistic backtesting, Monte Carlo, and prop-firm risk modelling. A niche, but a deep one — MNQ and MGC futures specifically. If you want someone who's read the fills data carefully, this is that.

  • Pine Script strategy review and rewrite
  • Walk-forward and regime-aware testing
  • Realistic fill modelling and commission drag
  • TopstepX account monitoring (read-only)
Shipped, not slideware

A partial catalogue.

Hotel AI
Belliq Sentinel. An exception-monitoring layer for hotel operations. 225+ tests · MCP server · voice + booking flow
RAG infra
Vector Brain. A 15-tool MCP server over Qdrant and Neo4j. 11 collections · 98,000+ chunks · hybrid search
Backtesting
Strategy Lab. 25 tools for honest, walk-forward backtesting. Monte Carlo · regime-aware · prop-firm risk modelling
Education
Pytor. Interactive Python tutor, client-side via Pyodide. 400 exercises · Groq-primary tutor bridge · free to use
Research
Nine-scout research pipeline. Reddit, GitHub, HN, arXiv, and more. Adaptive scoring · Groq 70B synthesis · daily digests
Visualization
The brain. Interactive 3D view of the whole workspace. R3F + Vite · 280+ neurons · live semantic layer
The honest version

Who you'd be working with.

No computer-science degree. No prior industry experience. Older than the median AI engineer, based in Norway, working from a Linux laptop with a preference for local-first infrastructure and tools I can actually reason about.

Over the last year I built a handful of interconnected systems — a hotel AI with full test coverage, a retrieval pipeline across eleven collections, a quant research engine, a Python tutor, a research swarm with adaptive scoring, and a visualization of it all — using Claude Code as the primary development environment. I understand the current AI tooling ecosystem from the inside. I know what works, what wastes time, and where the rough edges are.

What I bring to an engagement is taste for the simple, a stubborn insistence on tests, and a refusal to pitch work I haven't done. What I don't bring is a pre-packaged deliverable or a big-agency surface. If that's what you need, I'm not the right person.

Get in touch

Have a project in mind?

Send a short note: what you're building, where it's stuck, and roughly when you need it done. I'll reply within 48 hours with either an honest assessment or a referral to someone who fits better.

matswm86@yahoo.no Norwegian or English — whichever suits you.