Work
Four projects, four industries, one of them talked out of existence.
We don’t publish client names. Every contract we sign forbids it, and we think that’s right. Industry, scale and results we can share, and we’ll put you in touch with a reference once a conversation gets serious.
1,100 employees · claims operations
Health insurer
- Challenge
- 38,000 claims a month classified and routed by hand.
- What we built
- Multimodal extraction from the forms, classification verified against the policy, and a review queue for anything uncertain.
- What we learned
- The review queue turned out to matter more than the model. Sending the uncertain 36% to a person, with the extracted fields already filled in, is what made the operations team trust the other 64%.
- 64%
- routed with no human touch
- 9 → 2 days
- average resolution time
480,000 part catalog
Industrial distributor
- Challenge
- Internal search couldn't find parts by standard or by equivalence.
- What we built
- Hybrid retrieval — keyword plus vectors — with reranking and an industry equivalence dictionary.
- What we learned
- Vectors alone made it worse: they returned parts that were similar rather than compatible, which in this business is the difference between a sale and a return. Keyword search on standards codes carried the result.
- +31%
- conversion from search
- −46%
- calls to the technical desk
210 attorneys · vendor contracts
Corporate law firm
- Challenge
- First-pass contract review, repetitive and expensive.
- What we built
- Clause extraction under a strict schema, comparison against the internal playbook, and summaries citing the exact paragraph.
- What we learned
- Nobody would use it until every finding carried a paragraph citation. Attorneys do not accept an answer they cannot check in ten seconds, and they are right not to.
- 340 hrs
- of review saved per month
- 100%
- of findings with a verifiable citation
Underwriting operations
Consumer lending platform
- Challenge
- They asked us to build a model that scores credit applications.
- What we built
- We recommended against it: high regulatory exposure and historically biased data. We built incomplete-documentation detection instead.
- What we learned
- The project we talked them out of would have been larger and better paid. The one we built removed more friction from their funnel than a scoring model would have, and nobody has to defend it to a regulator.
- −52%
- applications returned for paperwork
- 0
- automated credit decisions
Yours would be the fifth.
Tell us the process that hurts, what systems hold the data, and the deadline you’re working to.