Work
Three case studies
Subject
AI for slow, manual work
Files
01 to 03

Three AI projects, start to finish, with the parts that went wrong left in.

Each one starts with people doing slow work by hand. I decided where AI should help and where a person stays in charge, built the first working version myself with AI, and tested it on real work.

  1. Case study 01Logistics

    Mapping Studio

    A 20-hour specialist job, drafted in 8 minutes.

    Measuredthe tool’s own timing
    Real screen
    The rebuilt tool: a four-step header, filter chips with counts, verdict tallies, and the field tree beside the selected field’s details, laid out like the analysts’ own software.
    The rebuilt tool, made to feel like the analysts’ own software. Client names blurred.

    Before a new customer can go live, an analyst matches every field of their data by hand. I watched one being done, then built a tool that drafts it and shows its evidence.

    What went wrong

    The first version’s output was good, and the analysts went quiet. I rebuilt it to feel like their own software, and they started relying on it.

    My part
    Discovery to delivery; I directed the build
    Status
    In production on the company’s own servers
  2. Case study 02Supply chain

    EDI Copilot

    A compliance check on customer data the AI never sees.

    Diagram
    1. The file, on the host

      Stays on the host

      N1STMEGA MART STORES

      The customer’s name, as it arrives.

    2. What the AI sees

      Sent, once a person approves

      N1ST[PARTY_NAME_1]

      A token in its place. The name never leaves.

    One line of an invoice file, before and after masking. The name is a test value, not a customer’s.

    Analysts check invoice files against a retailer’s rulebook, 30 to 50 minutes a file. The files carry protected customer data, so the AI reads only masked copies, and a person approves each one first.

    What went wrong

    I shipped a confidence score, and the pilot users couldn’t act on it. I deleted it, and two other controls with it.

    My part
    Every product, safety and evaluation decision
    Status
    Piloted by two to three analysts
  3. Case study 03Education

    Chalkline

    Lesson videos for a school, designed with seven points where a person signs off.

    Diagram
    Where the finished video’s twelve beats went.

    A school wanted its textbook lessons as short animated videos for young children. I designed a twelve-stage system that turns a lesson into a narrated 3D video.

    What went wrong

    Then I checked one finished video against the lesson. It taught none of the lesson’s exercises, and both checks had passed it.

    My part
    Every product, evaluation and budget decision
    Status
    Pro bono, for a private school in India