I run my businesses on the thing I sell.

Self-hosted AI operating substrate

One console that runs the whole operation — conversation with local and hosted models, a governor that decides what an agent is allowed to do, a fleet that runs jobs overnight, retrieval across my own documents, and a vision endpoint. It runs on hardware I own. Nothing about how my businesses work has to leave the building for it to function.

• Runs local models via Ollama, no external dependency for core function
• Governor layer gates what agents may touch — the part most AI deployments skip
• Written to run with essentially no third-party packages, so it doesn't rot

Private knowledge graph and timeline

Every decision, document and event across two businesses, linked into a navigable graph and a dated timeline. It answers questions like "why did we decide that, and what did we know at the time" — the question that normally has no answer once a business is more than a year old.

• ~499 linked nodes across the working knowledge base
• 60-day rolling timeline, rebuilt from source rather than hand-maintained
• Served locally only — the graph never leaves the machine

Search that actually knows the business

Roughly 2,300 indexed chunks across two document vaults, searchable by meaning rather than keyword, wired into the same console I work in. This is the piece most businesses think they're buying when they buy an AI tool, and it's the piece they usually don't get, because the vendor can't see their files and wouldn't be allowed to if it could.

• Indexes the real corpus — contracts, decisions, specs, correspondence
• Runs against local embeddings; sensitive material never goes to a hosted model
• Re-indexes on a schedule, so answers don't quietly go stale

Work that happens without me

A set of standing programs that run on their own schedule — research sweeps, monitoring, document preparation, morning briefings. Each one produces a draft for review. None of them send anything, publish anything, or spend anything. That constraint is the design, not a limitation.

• Multiple standing programs on schedule
• Draft-only by rule: nothing leaves without a human pressing send
• Output lands as reviewable artefacts, not as actions already taken

Pipeline and invoicing

A small, dependency-free CRM that tracks leads through to paid invoices, built because every hosted option wanted a monthly fee and a copy of my customer list to do arithmetic I could do myself.

• Leads → deals → invoices, on a command line
• Data stays in flat files I can read without the software

Monitoring across my own estate

Continuous integrity and network monitoring across my machines, with a tamper-evident record of what changed and when. I'm deliberately not publishing its architecture — describing your own defences in detail is how you stop having any.

• Tamper-evident change record
• Architecture intentionally undocumented in public

Web estate, self-deployed

Both properties are built and deployed from my own pipeline, with no agency in the loop. Unglamorous, but it means a change to positioning takes an afternoon rather than a change request and a fortnight.

• max-io.co and armatech-industries.com
• Own build and deploy path