Skip to content
Cameron Mills

Projects

Start with these three cases. Each page opens with the user, the problem, the options, the decision and what came of it, and the technical detail follows.

  1. Tearline

    A company name in, a cited tear sheet out

    User
    Analysts and lenders on PE, credit and restructuring desks who work on UK private companies.
    Problem
    They start from scanned accounts and build each tear sheet by hand, so the work is slow and the numbers are hard to trust.
    Decision
    Public-register data with page-level citations, trading data breadth for legal safety. Where printed figures do not reconcile, such as a maturity split that does not add to total borrowings, the engine abstains and says why. A sales line that promised absolute accuracy was cut, because nothing supports it.
    Outcome
    Open source on GitHub, and deployed behind its sign-in wall since 30 September. The checks so far are engineering checks: 1,362 passing tests, and an extraction evaluation against a gold set of 10 real UK filings with hand-checked figures, where abstentions are reported apart from failures. No field-level error rate or analyst time saving is published yet, and page pinning on scanned accounts has not been checked by a person.
    Stage
    Open source on GitHub

    Read the case: Tearline

  2. Life Tracker

    A personal AI agent system, now used by a beta group of 10

    User
    Me first, and now a beta group of 10 selected users. Its first job is making sure nothing falls through the cracks: deadlines, admin and application cutoffs.
    Problem
    Obligations arrive in calendars, mailboxes and voice notes, and keeping a tracker up to date takes the time it is meant to save.
    Decision
    Agents do the upkeep inside a safety layer: no agent on the Mac can make a payment, and anything an agent may not do comes to me as a one-line command. I set the model-routing rules from an analysis of 519 agent runs by task type and model. Those are runs, not users.
    Outcome
    Used by a beta group of 10 selected users. Acting on their feedback, I ran refactoring rounds to simplify the codebase and widened Chief of Staff's predictions to cover more tasks. This page does not yet define active use, give a feedback example or show a measurement taken after those changes.
    Stage
    In beta with a group of 10 selected users

    Read the case: The Life Tracker

  3. Emerald Consulting OS

    Replacing a no-code workspace with an owned, private system

    User
    The 5-person leadership team of INSEAD's pro bono consultancy, who run recruiting, clients and offers, and the applicants and clients who use its public pages.
    Problem
    The club ran on a monday.com workspace full of automations, but it holds applicants' CVs and clients' details, so it needs to own that data and control exactly who sees what. The aim was one system for the whole year, with no side spreadsheets.
    Decision
    Build the club its own system. One concrete choice: a reviewer sees a candidate's name, contact details and CV only after saving their own scores, so first reviews are fair.
    Outcome
    The public pages are live and the leadership team runs the club on it. On my CV I report that, against prior years, client relationship retention rose 40% and the share converting into projects rose 20%. The counts and comparison periods behind those figures are in the club's records and are not shown here, so this page does not claim the software caused all of the rise.
    Stage
    Public pages live

    Read the case: Emerald Consulting OS

Other main builds

  • A nightly agent that plans ahead, with a logged undo and one-tap approvals

    Chief of Staff

    A nightly agent that plans ahead, does reversible work with a logged undo, and turns the rest into one-tap approvals. After beta feedback its predictions were widened to cover more tasks. Relaunching the mail client is the one action with no undo, and the public core in the repository makes no model calls.

    Result: All three ship switched off behind a flag file, and the chief of staff has run live since 14 October

    Role
    Designed and directed. Code and checks written by Claude Code agents to my specification.
    Status
    Chief of staff live since 14 October, follow-on still switched off
    • Python
    • launchd
    • Claude Opus
    Code on GitHub: Chief of Staff
  • A video and design engine that turns sources, a script and a voice into a finished short

    Digital media system

    Sources, script, voice and timing drive a beat-aware edit engine that makes the short films.

    Result: One engine now makes the Green Label weekly briefing video and the Fontainebleau ad

    Role
    Designed and directed. Code and checks written by Claude Code agents to my specification.
    Status
    In use, with an explainer film made by the engine
    • Python
    • ffmpeg
    • Pillow
The full catalogue of parts

Each is a part of one of the main builds, listed under its parent. They are engineering detail rather than products of their own.

Part of Life Tracker

Part of Chief of Staff

  • Model substitutionFinding the AI calls that do not need a frontier model
  • Drift testingTesting whether browser agents still work when the web changes
  • nightly noteA nightly note on what to do next, from one Opus call
  • Chief of staff actionsFive safe actions the chief of staff may take on its own
  • Follow-onThe next steps after a run finishes, queued while the context is fresh
  • strategistA Sunday deep dive that proposes builds and never starts them
  • Action log and undoOne log of every automatic action, with its undo written first
  • Polaris surfacesPolaris screens for reading each system's dry run and its actions
  • Approval flagsA flag file per system that keeps it off until I approve it

Part of Digital media system

  • SourcesFeeds and pages read into a ledger where every number is checked
  • Green Label content engineOne sourced spec in, a newsletter, a post and a short video out
  • VoiceNarration on my own machine, with room for a recorded voice
  • TimingWord times from the voice and a beat grid from the music
  • Design libraryA library of learned edit styles and the engine that applies them
  • Creative appsOpen-source photo, vector, layout, video and motion apps my agents can drive
  • Captions and exportCaptions drawn frame by frame, then one encode and a self-check
  • Fontainebleau adAn ad cut to a reference clip's beats, with every sound synthesised
  • One Paper FurtherA short film that explains one AI research paper, and where it stops