Joel Stevick

Blog · RSS

Built, then measured

Engineering write-ups on building with AI, every number from a run committed beside it.

  1. EvalsAsk the book

    What a golden set found in my book's chatbot

    Thirty-two questions, three runs and under two dollars. The book's index had never built, code questions never got their code, and my scorer was older than my rules.

  2. Tuning the demo

    The typewriter effect, with a fact-checker in the way

    The demo checks every answer against its sources, so I taught it to type one checked sentence at a time. The eval set caught a version losing 8 of 26 answers.

  3. Running the shop

    Teaching my site to run the shop

    Every five minutes my site works out what is due and does each thing it can, as me. It tells me only what needs my hands, one line with one button.

  4. Finding clients

    Teaching my site to find its own prospects

    One button on my admin page searches the web for companies that may need my services, checks each find in code, and learns from earlier finds.

  5. Tuning the demo

    Giving the answer the rest of the point

    Adding each passage's neighbours answered all 5 two-passage questions, up from 4, at 60% more a question, over my 50% limit. I switched it on anyway.

  6. Tuning the demo

    “The passages don't list his job history.”

    My demo chatbot said my résumé listed no jobs. How I fixed it test-first, then tuned the pipeline with a 25-question eval on live models, for $1.96.