I was supposed to plug the Claude API into my chatbot. Instead I built my own AI engine from scratch.
Monday's plan had four tasks — simple, realistic. Here's what actually happened across six retail shifts and one Wednesday off.
Planned → shipped
- Dashboard buttons redesigned — dark, bordeaux, 7 vector icons drawn on Canvas. The main screen finally stopped looking like a prototype.
- Auto RAM flush at >75% — the first TURBO feature that actually intervenes: detects sustained pressure, flushes.
- GuideAI Google OAuth fixed — login works again. Sounds small, but nothing else can happen until users can sign in.
- GuideAI Claude API — not yet. Still researching how to secure the API key in a public project. I'm not shipping something half-baked that leaks credentials.
Unplanned — and shipped anyway
A new hybrid AI engine for hck_GPT. Not an LLM — a rule-based command parser that actually tries to understand what the user is asking. More commands, better intent detection, built to be the default AI core. Ollama integration comes around v1.8–1.9 for anyone who wants a real local LLM (heavy in the background, so it'll be opt-in — PC Workman itself stays under 1% CPU).
Why build this instead of just plugging in Claude? Because not every user has an API key, and not everyone wants cloud AI. The app needs to be smart on its own first. External AI is the upgrade, not the foundation.
The hck_GPT chat banner also got completely rebuilt in the dark bordeaux style. I don't know how long it looked that rough — months, probably. Nobody told me. I didn't notice until this week.
Three of four planned tasks shipped, plus an entire AI engine nobody asked for. Sometimes the best features are the ones you didn't plan. v1.7.2 .exe releases this weekend.