HOWO Parts Decision Engine
An evidence-first backend that turns rough Telegram messages, photos, and voice notes into structured, reviewable spare-parts knowledge.
Capture → validation → verified result
A sanitized test through the real decision pipeline.
The 1.0 score is the engine’s score for one exact-match test, not an accuracy rate. This is a working prototype, not a finished production inventory application.
The problem
Parts knowledge arrived in forms a normal database could not trust.
Similar components, inconsistent naming, workshop language, and incomplete photos make an unverified AI answer financially dangerous.
The system
AI proposes. Deterministic logic decides what is trusted.
The workflow preserves raw evidence, extracts a structured proposal, validates the schema, ranks known products, and routes ambiguity to a person instead of silently confirming a guess.
An exact known-part match without unnecessary human review.
The test passed validation, matched the known part number at confidence 1.0, wrote the observation and captured part number, and created no review item.