Analyze

Retrieval — what the chatbot is told before it answers

The chatbot used to receive the entire knowledge base and the entire rate card on every message. Now it receives only the facts the question is about, retrieved from an index of three sources: the company knowledge, every job on the rate card (with the symptom words customers actually type), and what has really been reported. Measured at a 68% smaller prompt with a realistic knowledge base — and the gap widens with every fact you write down.

Token savings — how much the brain answers itself

Every photo is answered one of three ways. Dedup reuses the verdict of a near-identical photo, own model answers from what it has learned, and cloud sends it to a paid engine. Only the last one costs money, so this is the number to drive down — and it is measured from what was actually recorded, not from a counter kept alongside it.

Chat photos — confirm or correct the AI

Every photo a customer has sent, newest first, with what they wrote alongside it and what the model concluded. Confirming or correcting one teaches the model straight away: the exact job goes into the confirmation memory for instant precision, and the category becomes a weighted training sample — a correction counts for more than another easy example, because it teaches the model what it got wrong. The queue below only shows what the brain was unsure about; this shows everything.

Review queue — active learning

The brain flags photos it was unsure about, or where it disagreed with the AI that answered. Confirm the real type here — these are the highest-value labels, and they feed straight back into training as weighted corrections.

Autonomy — self-improving, safely

The model retrains itself as new labels arrive, but a new version only goes live if it beats the current champion on a frozen golden test set — so it can run unattended and never regress. Rejected challengers are kept for history, not deployed.

Our model — text · card & training

ML diagnostics — where it's confident vs confused

Cross-validated confusion matrix for the intent model (rows = true intent, columns = predicted). Shows which questions get mixed up — same analysis as the vision model.

Self-learning & autonomy — the chatbot teaches itself

When an LLM answers a question the model didn't know, it remembers the Q&A (and matches future paraphrases). It re-studies as it learns, and a new version only goes live if it beats the champion on a frozen golden set — the same safety gate as vision.