Most AI assistants have one layer: send everything to ChatGPT. Mine has nine. And I didn't plan any of them.
Monday's post hit 13,500 views — people saw the app. Today I'm showing how it thinks. Every message typed into hck_GPT goes through a 9-step decision waterfall before it gets a response. Not because I designed it that way — every layer exists to fix a problem the previous one couldn't handle.
The routing waterfall, stripped to the decisions
- Service wizard active? The wizard handles it.
- Hard-route keyword? Bypass AI entirely.
- Help keyword? Instant help card.
- "Why" detected? Special why-handler, before the parser.
- Single-word alias ("cpu", "ram")? Instant info, no parsing.
- AI intent parser → confidence score. High confidence → rule engine (fast, deterministic). Low confidence + Ollama → local LLM (smart, contextual).
- Entity found but no intent? Route by entity type.
- Context carry-over — "and?" repeats the last topic.
- Nothing matched — graceful fallback.
Type cpu and it never reaches the parser — instant alias. Type why is my RAM high? and the why-detector catches it at layer 4. Type what should I do about my PC being slow? and the parser returns low confidence, so Ollama gets it with full system context.
The part I had most fun building: intent-aware temperature
Every AI tutorial uses one temperature setting. I use 20.
- "What CPU do I have?" — temperature 0.35. I want a surgeon. Exact facts.
- "How are you?" — temperature 0.80. I want a mate. Warm, funny.
- "Why is my PC slow?" — temperature 0.45. Diagnostic but readable.
The whole engine in three lines:
confidence >= 0.60 -> rule engine (instant, deterministic)
confidence < 0.60 -> Ollama LLM (smart, full PC context)
Ollama timeout -> rule fallback (best effort, never crashes)I didn't want Ollama for everything — it's slow. I didn't want if/else for everything — it's stupid. So I gave every message a confidence score and let the score decide. Nine layers. None of them planned. All of them necessary.
Next week I start Umiejętności Jutra 3.0 — Google's AI certification through SGH. Can't wait to see how it changes the way I think about this engine.
Does a technical format like this, once a week, land for you — or is it too in-the-weeds?