OPERATIONS · TRIAGEB2B SAAS · CLAIMS OPERATIONSPLACEHOLDER CASE · RIGHTS PENDING
Nuforce —
Claims triage engine.
73% auto-resolve.
A direct-LLM triage layer over Nuforce's legacy claims pipeline. Reduced human review by 73 percent in twelve weeks. No agent framework, full ownership at handoff.
73%
Tickets auto-resolved at confidence ≥0.85
412 ms
Median end-to-end latency
98.2%
Eval pass rate on holdout set
8 wk
Kickoff → production traffic
01 · THE PROBLEM
Where they were stuck.
Nuforce was hiring claims analysts faster than it could train them. Inbound volume had tripled in eighteen months, and a queue of 8,000 unresolved tickets had built up. The team had already trialled two off-the-shelf 'AI triage' tools — both required massive prompt-engineering retainers and neither delivered measurable deflection.
02 · OUR APPROACH
How we built it.
- 01Two-week scope sprint to label 500 representative tickets across 12 intent classes
- 02Direct LLM classification pipeline (no framework) with structured-output schema
- 03Confidence thresholds — high-confidence tickets auto-resolve, mid-confidence enter a human-in-the-loop queue, low-confidence escalate to a senior analyst
- 04Eval suite running on a holdout of 1,200 historical tickets, gating every deploy
- 05Cloudflare Workers + Postgres + Notion review queue — boring infrastructure
“We expected a slow rollout and a long tail of manual exceptions. We got the inverse — clean deflection from week three, and our analysts now spend their time on the cases that actually need a human.”
Engineering Lead
Nuforce
[TECH STACK]
- Postgres
- Python
- Cloudflare Workers
- Notion API
- OpenAI
[ENGAGEMENT]
Duration12 weeks
ShapeOPERATIONS · TRIAGE
HandoffFull ownership
Warranty90 days