Skip to content

Builder project · Built at Pulse Foundry AI NYC X ABI Frameworks

Concur

confidence comes from sources concurring

Share on LinkedIn

Tip: click @Pulse AI in the post to tag us.

From idea to reality

What we built

# Concur **Turns the highest-denial paperwork problem in skilled-nursing into a one-glance accept/flag/reject decision.** ### The problem Wound-care claims get denied **25–35%** of the time — almost always over **one missing measurement** (usually depth) nobody catches until the claim bounces. Today a biller reads hundreds of charts across **4 messy note formats** by hand and guesses. ### What we built A pipeline that pulls the chart data and tells the biller, per patient: > **`FA-001 — FLAG — Depth not documented; only the note describes the wound. Confirm before billing.`** …with the source note highlighted in place and a graph of which sources agree. ### The core idea: **flag, don't hallucinate** - **Regex owns every number** — a measurement is only ever a *literal substring* of the note. The system *cannot* invent a depth. - **Confidence = cross-source agreement**, not the model's word. High confidence only when the diagnosis, note, and assessment describe the *same* wound. ### How it works (7-stage resumable pipeline) | Hard problem | Solved by | |---|---| | API fails **30%** of calls (429) | backoff + anti-storm semaphore + checkpointed resume — **0 lost records** | | One patient, **two IDs** | resolved once in a hard gate → wrong-key errors impossible | | 4 note formats | detected from text; regex measures, optional Claude reads prose (span-gated) | | Auditable decision | routing is a **SQL view** with a Python oracle CI proves equivalent | **Live: 300 patients · 556×429 absorbed · → 12 auto-accept · 92 flag · 196 reject**, each with a plain-English reason. ### Why it's real Versioned migrations, idempotent, resumable. Hardened for production: incomplete-fetch charts **never auto-accept**, incremental re-runs (300 skipped in 0.13s), cached+concurrent LLM. **The whole pipeline runs with zero model calls** — Claude is upside, never a dependency. **One line:** *Pulls messy two-keyed EHR data through a failing API without losing a record, extracts wound facts from four formats without inventing a number, scores trust by source agreement, and hands the biller a decision they can defend.*