We embed engineers inside carriers to evaluate, clean up, and build the AI infrastructure behind claims intake, fraud flagging, and retrieval — held to ground-truth benchmarks, not vendor slides.
Most carriers have a claims AI demo that impressed a steering committee and then stalled. We turn that into infrastructure that adjusters trust and compliance can sign off on — across three surfaces.
Agents that read FNOLs, forms, and supporting docs, extract structured fields, and route submissions — with a human-in-the-loop where the stakes demand it.
Retrieval-grounded models that surface duplicate claims, inconsistent narratives, and network-level collusion — with precision the SIU actually wants to action.
The unglamorous layer that decides whether the whole system works: chunking, embeddings, and retrieval that actually surface the right policy clause and prior claim.
The difference between a claims AI that works and one that quietly leaks errors is decided at four layers. This is where our Forward Deployed Engineers spend their time.
Claim corpora are forms, policy schedules, adjuster notes, and medical reports — not blog posts. Fixed-size splits shred tables and separate a clause from its exclusions. We chunk to structure.
A claim query is never "find similar text" — it's "find similar text for this policy, this loss type, in force on this date." Metadata filtering and hybrid search are non-negotiable.
Retrieval-Augmented Generation for the long tail of claims and policies. Cache-Augmented Generation for the stable, high-volume schemas — preloaded context, lower latency, fewer moving parts to break.
If it isn't measured against adjudicated ground truth, it isn't shipped. We build the golden datasets and the harness, then wire the thresholds into CI so a regression blocks the merge — not the quarterly review.
| Metric | Score | Gate | Status |
|---|---|---|---|
| Retrieval recall@8 | 0.91 | ≥ 0.85 | pass |
| Answer faithfulness | 0.96 | ≥ 0.95 | pass |
| Hallucination rate | 0.014 | ≤ 0.02 | pass |
| Fraud precision | 0.88 | ≥ 0.80 | pass |
| Fraud recall | 0.94 | ≥ 0.90 | gate |
| Extraction F1 | 0.93 | ≥ 0.90 | pass |
| Latency p95 | 0.82s | ≤ 1.0s | pass |
| Cost / 1k claims | $3.10 | ≤ $4.00 | watch |
Vendors show you a highlight reel. We show you the scorecard — built on your own adjudicated claims, run on every change, wired to block a release that regresses.
Not a slide deck and a Slack channel. Senior engineers who sit with your claims and data teams, work in your repos and your data environment, and leave behind infrastructure your people own.
Map the current claims stack, data, and the demo that stalled. Find where trust breaks.
Week 1–2Build the golden dataset and harness. Get an honest baseline of what actually works.
Week 2–4Fix chunking, retrieval, and agents against the evals. Ship the first gated release.
Week 4–10Drift monitoring, CI gates, fraud thresholds, on-call runbooks. Production-grade.
Week 10–16Your team owns and extends it. We're a reference call, not a dependency.
OngoingTell us where it's stuck. We'll scope a Forward Deployed engagement and show you the evals it would have to pass.