Writing · 4

Writing

Notes from building and running production systems: what the problem was, what I built and what went wrong.

  1. Letting an AI assistant query school records without showing them to the model

    How ByteLift's assistant answers from role-specific manuals and looks up live class data, while the language model only picks a tool and never reads a student's result.

  2. Replacing a hosted chatbot with RAG on local models

    How SkinStation's website assistant answers from approved documents on a single GPU, refuses when it has no evidence, and hands the conversation to a person.

  3. Safe deploys on one small server: staging, backups and automatic rollback

    How this site ships from a git push to production on a single server, with a rehearsal, a backup, a staged candidate and a rollback that has already had to work.

  4. Starting laundry machines from the cloud: four versions of one command

    How Quicklean's washers and dryers went from a relay computer in every branch to chips that connect directly, confirm each start and report a heartbeat.