AI Solution · 2024
Atlas Copilot for Atlas Logistics
A retrieval copilot over two decades of freight contracts, evaluated on every release.
- 94%
- Eval accuracy
- 5/wk
- Hours saved / analyst
- 180k
- Documents indexed
The challenge
Contract analysts spent roughly six hours a week locating clauses across 180,000 scanned documents, and confidence in generic AI search was low after a failed pilot.
What we did
Built an OCR and chunking pipeline tuned for scanned freight paperwork.
Combined hybrid retrieval with reranking, then grounded every answer in a citation.
Shipped an eval harness of 400 analyst-written questions gating each deploy.
Capped per-query cost and latency with a routing layer across model tiers.
The outcome
Analysts recovered roughly five hours each week, and answer accuracy holds at 94% against the eval suite release over release.