Governed AI agents, deployed inside real businesses. Embedded until they actually work.
That's me
One of Integrate.AI's product offerings: a multi-tenant SaaS platform for governed AI agents. Tenancy enforced at the database with Postgres row-level security, Stripe billing reconciled inside tenant-scoped transactions, every agent action written to an audit spine.
first client onboarded: a multi-site hospitality groupA firm-owned memory layer, currently in stealth: agents that wake up already knowing the people, the projects and the decisions. Always-loaded core context plus an unbounded archive recalled through hybrid vector search, with a reflection loop that distils sessions into durable facts overnight. It also grades its own health every day.
live daily as the memory spine of my own agent fleet
A native macOS dictation app I built to replace a paid subscription. Speech to text runs locally on the Neural Engine, then Claude Haiku tidies the transcript before it pastes. If the cleanup ever fails, the raw words still paste, so dictation never breaks.
signed, auto-updating, in daily use
Eight self-contained skills that cut what an agent pays to read PDFs, spreadsheets, video, logs and web pages. They plug into any Anthropic API harness, and every savings claim is produced by a meter against a fixture corpus and committed as evidence.
every number links to its meter report in the repo View code ↗For workloads where a frontier API is the wrong answer: LoRA fine-tuning Qwen3-4B on a 16GB MacBook Air with Apple MLX, mapping macro snapshots to tactical asset-allocation tilts entirely on device. Ships with a research brief on methods, costs, and the break-even against frontier APIs.
about £0 marginal training cost on device View code ↗A worked sample of graph engineering: an autonomous daily macro briefing in an asset manager's house style, orchestrated with the Claude Agent SDK in Python. A manager sets the day's focus, three specialist sub-agents research in parallel with live market-data tools, and a verifier scores every memo against a nine-point style checklist. Failures get one retry with feedback.
memos gated at 7 of 9 on the style checklist View code ↗Deployed for clients of Integrate.AI and anonymised here. Named references on request.
After a sales call, it picks the right contract and order forms, fills them in, files them and drafts the follow-up email with everything attached. The rep reviews and sends. It never sends on its own.
A marketing engine for a multi-site wine bar group: plans, drafts and schedules campaigns across channels from one brief.
Turns enquiry-to-event admin into structured function sheets the operations team runs the venue from.
A daily personal intelligence brief covering markets, mail and diary, rendered before the client wakes.
The firm's own entities and relationships ground every agent answer and gate every agent action. The model is never the source of truth about the business.
The more an agent knows about how the business actually runs, the better its work gets. Most of that knowledge sits in people's heads, so the first job on every engagement is getting it out of their heads and onto paper.
Some of those tasks can be automated, which hands back the hours for the part of the job that is uniquely yours, and usually the most interesting part. This is not about replacing people. I am not in the business of asking turkeys to vote for Christmas.
Every side-effect routes through a tiered autonomy gate: auto, draft-then-notify, or confirm. Enforced in the runtime, written to an append-only audit log. Not a prompt.
Savings get meter reports. Self-modifications pass held-out eval gates. Style gets a scored checklist. If a number is not measured against committed evidence, it does not ship.
Consulting decks do not change businesses; deployed workflows do. I sit inside the client's operation, their inbox, their events, their expenses, until the system is part of how they work.