A vehicle-data platform built to refuse guessing
We built a vehicle-data enrichment platform that queries two language models, cross-checks their answers, and logs every disagreement - instead of quietly filling gaps with a plausible-sounding guess.
Enrichment that doesn't know when to stop
LLMs are good at filling in missing vehicle specs - and just as good at confidently inventing specs that don't exist. For a database other systems depend on, a plausible-sounding wrong answer is worse than no answer at all.
The goal was a pipeline that enriches real gaps in vehicle data while refusing outright to extrapolate or guess at anything it can't verify.
Enrichment with a hard no-guess boundary
Four stages turn a known make, model, and year into a verified, schema-valid record - or into nothing at all.
Structured query Scoped input
Every enrichment request is built from a known make, model, and year, and sent as a tightly scoped structured query - not an open-ended prompt a model can wander off from.
Dual-model cross-check GPT-5 + Gemini
The same query goes to two different models, and their answers are compared against each other rather than trusted individually - a single model's confident mistake doesn't make it into the database unnoticed.
Schema validation Strict JSON
Every response has to satisfy a strict JSON schema before it's accepted. Malformed, incomplete, or out-of-scope answers are rejected outright, before they ever reach the database.
Conflict logging & no-guess boundary US-only
When the two models disagree, the conflict is logged rather than silently resolved. The pipeline is scoped to US-only data, with a hard rule against extrapolating beyond what it can verify.
Built to be wrong loudly, not quietly
What it delivers
Engineering behind it
Need data enrichment you can actually trust?
We build pipelines that know the difference between filling a real gap and making one up - with every disagreement logged, not hidden.
Book a discovery callInternal Waveo system, described by architecture and outcome.