B2B lending platform · 11 weeks
A lending team read every application pack by hand. We built an extraction and classification workflow with a human review queue, benchmarked on two years of their own files.
Median review time, from 3 days
Extraction accuracy on live packs
Applications processed per analyst
Each application arrived as a pack of statements, filings, and identity documents in inconsistent formats. Analysts read them end to end, and turnaround time was costing the business deals.
Assembled an evaluation set from two years of historical applications with known outcomes before writing any production code.
Built extraction and classification for each document type, with confidence thresholds tuned against that set.
Routed low-confidence and high-value cases to an analyst review queue with the source passage highlighted.
Logged every model call and decision for audit, with accuracy tracked continuously after launch.
Analysts now review exceptions rather than reading every page. Median turnaround dropped from three days to under an hour, and the audit trail satisfied the compliance review on the first pass.
“They spent the first three weeks proving it would work on our worst files. That is why we trusted the rest of it.”
More Work
Logistics
A third-party logistics operator was rekeying every customer order from email and portals into its warehouse system. We replaced the whole path with a monitored ingestion pipeline.
Healthcare
Eleven clinics each ran their own rota spreadsheet. We built a single scheduling platform with the compliance rules encoded, not remembered.
Get In Touch
