Projects

Each card is a summary. Open one for the problem, the approach, the architecture, what I built and what came of it.

The Life Tracker agent system
Running unattended AI agents reliably
A manager and a supervisor for headless Claude Code agents, so runs left alone overnight either finish properly or stop in a way I can see and resume.
249 runs recorded done, 3 killed, 2 timed out (15 to 29 Sep 2026) (agent-reported)
Role
Designed and directed. I run it every day. Code and checks written by Claude Code agents to my specification.
Status
In daily use, scheduler running every minute (repo at c909757, 1 Oct 2026)
- Python
- Claude Code
- Swift
- iCloud Drive
Objective Mind
Shipping a live AI product end to end
A founder-accountability web app for small pods of founders, taken from HTML prototypes to a hosted product with sign-in, sync and AI briefings.
Shipped at objectivemind.co.uk in July 2026 (agent-reported)
Role
Founded, designed and directed. Shipped at objectivemind.co.uk in July 2026.
Status
Shipped at objectivemind.co.uk in July 2026 (handoff dated 20 Jul 2026)
- Next.js 16
- React 19
- TypeScript
- Tailwind v4
- Dexie
- Supabase
- Claude API
- Vercel
Tearline
Every figure cites its page
A company name or number goes in, and a tear sheet, capital structure, Excel model and references come out, with every figure tied to the document and page it came from.
17 of 18 critical-path work packages merged, 27 Sep 2026 (agent-reported)
Role
Co-founded. I designed the engine and directed its build with Claude Code. Pre-launch.
Status
Pre-launch, as of 27 Sep 2026 (launch goal check)
- TypeScript
- Next.js
- React 19
- Postgres
- exceljs
- Playwright
- Vercel (London region)
Polaris
One local dashboard where AI agents hand work back to me
A desktop dashboard that turns what my AI agents need from me into cards I can act on, such as questions to answer, commands to copy and run, and revision drills, served by a server only my own Mac can reach.
Commands the agents hand me are re-checked every 10 minutes, and a stale one loses its Copy button (agent-reported)
Role
Designed and directed. Code and checks written by Claude Code agents to my specification.
Status
In use, revision sessions run from it on 1 Oct 2026 (repo at c8e90f7)
- Python
- JavaScript
- WebGL2
- xterm.js
- pywebview
- tmux
Dreaming
An overnight AI pass over a day's notes, capped and checked
A nightly job that asks a small AI model whether pairs of the day's notes are related, duplicated or in conflict, and leaves a short report by morning. The model has no tools and a hard cap, and any answer that does not quote the notes exactly is thrown away.
Replayed against 90 past closes, the overnight close rule would have made 0, so v1 reports rather than edits (agent-reported)
Role
Designed and directed. Code and checks written by Claude Code agents to my specification.
Status
In progress, dry-run week, night 1 of 7 ran at 04:30 on 1 Oct 2026 with note writes off (repo at c8e90f7)
- Python
- Claude Code
- Claude Haiku
- launchd
- git
The knowledge bank and /initiate
Giving AI agents a memory that survives a cleared session
A cited knowledge bank that every unattended agent draws on, and a handover routine that lets a cleared Claude Code session pick up where the last one stopped, from written state rather than memory.
90-page bank installed with 724 of 724 quoted claims checked valid (30 Sep 2026) (agent-reported)
Role
Designed and directed. Code and checks written by Claude Code agents to my specification.
Status
In use, queued agent runs get its launch block (repo at c8e90f7, 1 Oct 2026)
- Python
- SQLite FTS5
- Claude Code hooks
- Markdown
- git
Steward, the managed agent runner
Keeping long AI agent runs alive through usage limits and full contexts
The runner behind every unattended Claude Code job. It winds a run down before the usage limit, resumes the same conversation after the reset, rotates a run whose context is too long into a fresh session, and feeds it all from a timed work queue.
Context rotation at 130k tokens, simulated to save 12% of agent spend (agent-reported)
Role
Designed and directed. Code and checks written by Claude Code agents to my specification.
Status
In use, every managed run goes through it (repo at c8e90f7, 1 Oct 2026)
- Python
- Bash
- launchd
- Claude Code
The /build skill
One build request in, a supervised multi-stage build out
Type one request and get a staged pipeline of unattended agent runs (discovery, research with adversarial review, a plan with acceptance tests, build, ship, final review), where each step opens only when a supervisor has checked the last one's work.
95 stages already passed in seven live pipelines were re-judged under the stricter 30 Sep checks, and all 95 still passed (agent-reported)
Role
Designed and directed. Code and checks written by Claude Code agents to my specification.
Status
In use, builds have run through it since 29 Sep (repo at c8e90f7, 1 Oct 2026)
- Python
- Claude Code
- launchd
- git worktrees
The Life Tracker voice layer
Turning speech into filed work, with the source of every transcript on record
Voice notes from an iCloud drop folder, a call recorder that transcribes on the Mac, and Wispr Flow added as a voice layer, all feeding one system that says what it did with each note and where each transcript came from.
Recorder self-test of 84 checks, with six cases on Wispr Flow, passing at merge (agent-reported)
Role
Designed and directed. Code and checks written by Claude Code agents to my specification.
Status
Live, Wispr Flow layer merged with the recorder and revise self-tests passing (repo at c8e90f7, 1 Oct 2026)
- Python
- Swift
- Core Audio
- Apple Speech
- Whisper
- Wispr Flow
LT Capture for iOS
An iPhone voice recorder that shows what the Mac did with each note
A SwiftUI app that records a voice note, hands it to the life tracker through iCloud Drive in a fixed file protocol, and shows each note's real status read back from the Mac's receipt.
Merged only after the five-step test suite passed on one commit; device checks still to run (agent-reported)
Role
Designed and directed. Code and checks written by Claude Code agents to my specification.
Status
In progress, built and tested, not yet installed on a phone (repo at 7acb911, 1 Oct 2026)
- Swift
- SwiftUI
- AVFoundation
- App Intents
- iCloud Drive
The humaniser and writing-style system
A house style check that will not let a rewrite drop a fact
A scorer, a written house guide and a review pass that every agent report goes through, measured against my own writing, with a compare step that fails any rewrite that loses a number, name or citation.
On 15 agent texts, em dash and semicolon-list lines fell from 19 to 2 under the guide (52 for the control), but 2 of 5 pre-set criteria were met (experiment)
Role
Designed and directed. Code and checks written by Claude Code agents to my specification.
Status
In use, every managed agent run carries the house guide and ends with the style review (repo at c8e90f7, 1 Oct 2026)
- Python
- Claude Code
- Markdown
- LaTeX
The revision engine
Spaced practice sessions that save every answer before the next one
A question-by-question practice engine that follows up until understanding breaks, keeps a weak-spot queue, and saves each answer to a sourced tracker before marking the next, in a chat session or on my dashboard.
Question 1 now comes from a pool written ahead, with no model call (it waited 30.7 s before) (agent-reported)
Role
Designed and directed. I use it myself. Code and checks written by Claude Code agents to my specification.
Status
In use, sessions run on the dashboard on 1 Oct 2026 (repo at c8e90f7)
- Python
- Claude Code
- Claude Sonnet
- Markdown