Project 15 - Sep 2026

Laya × RCM — Claim Denial Triage

Training Laya, my healthcare model, to triage insurance claim denials for revenue-cycle teams.

US providers lose an estimated $260B+ a year to denied claims, and most denials are recoverable — they're just never worked in time. Laya is the healthcare model I train and fine-tune on Hugging Face from open-web clinical and billing data. This side project points it at Revenue Cycle Management (RCM): read a denial, classify the reason, predict whether an appeal is worth it, and draft the appeal letter.

🧪 In progress — actively training

🤗 Laya on Hugging Face Follow the build on GitHub

Tech Stack

Laya (Hugging Face)

My fine-tuned healthcare model

Transformers + PEFT

LoRA fine-tuning pipeline

vLLM

High-throughput batch inference

RCM Data (835/EOB)

Denial codes + payer policies

The Problem

Denial reasons hide in free text and CARC/RARC codes that vary by payer
Billing teams triage by hand — the median denial sits for weeks before anyone touches it
Appeal deadlines (often 30-90 days) expire while the queue is being sorted
General LLMs don't know payer-specific rules and hallucinate policy language

The Idea

Laya fine-tuned on denial letters, EOBs and payer policy text gathered from the open web
Three heads: classify the denial reason, score appeal-worthiness, draft the appeal with cited policy
Served with vLLM for batch triage of a whole day's denials in minutes
Human-in-the-loop by design — a biller approves every letter; the model never files
Evals built around recovered dollars per hour of biller time, not just accuracy

The Triage Loop

Ingest — 835/EOB denial lines + letters normalised into one schema
Classify — Laya tags the root cause (eligibility, coding, authorization, timely filing, medical necessity…)
Score — predicts overturn probability × claim value → a ranked work queue
Draft — appeal letter generated with the payer's own policy language and the chart evidence needed
Review — biller edits/approves; every correction becomes new training data
Learn — periodic fine-tune on approved appeals; the queue gets smarter every week

By the Numbers

$260B+ annual US claim denial exposure
~65% of denials never appealed (industry estimate)
3 model heads: classify · score · draft

Open-web healthcare + billing training data
vLLM batch serving
Side project — actively training

Follow Laya

This is a work in progress — I post training runs and evals as they land. Estimates are for research; nothing here is billing or legal advice.

🤗 Laya on Hugging Face Follow the build on GitHub