AI for IME Report Writing

Workers' comp claims depend on an Independent Medical Examination (IME).

When someone gets injured at work, an independent physician reviews the entire case and writes a report on what happened, how severe it is, and what treatment makes sense. It carries real legal and financial weight, and it has to hold up if it's challenged in court.

~70%
of IME work is manual,
not medical
The goal

Take the manual work out of the process without taking the physician out of a single decision.

01
Obstacles

Volume

A single case can involve dozens of PDFs spanning hundreds of pages of treatment history.

Fact extraction

Conditions, medications, ICD-10 codes, dosage changes, test results, timelines - all buried across documents.

Citation tracking

Every claim must trace back to source document, page, and passage. Physicians track this mentally or with sticky notes.

HIPAA

Most AI platforms cannot legally handle PHI. The shortcut tools are off the table from day one.

02
Decision

Every project starts with one question:
buy, automate, build, or wait?

Buy

We find the software that already solves it, vet it against how you work, and get your team running on it. Live in days.

Automate

We connect the tools you already use so the manual steps disappear and data moves on its own.

Build

We design and develop something custom from the ground up. Built for how your business runs, owned by you.

Wait

We tell you when holding is smarter. You leave with a clear reason and a date to revisit.

Why We Choose Build

  • Buy failed: no off-the-shelf platform handles medical-legal citation requirements.
  • Automate failed: clinical judgment cannot be removed.
  • Wait failed: the pain is real today.
Audit Results

Before we recommend anything, we score the project on three things: how unique the work is, what a mistake costs, and how much of it runs on human judgment.

Company Specificity4 / 5
Off-the-shelf fitsOne of a kind
Cost of Error5 / 5
Low stakesMistakes are critical
Judgment Required5 / 5
Rote and rules-basedHeavy human judgment
03
How It Works

The AI extracts; the physician decides.

1

Upload

S3 + Cognito

Case files upload straight to a HIPAA-compliant S3 bucket using temporary Cognito credentials. The server never touches the file bytes.

2

Extract

Textract + Comprehend Medical

OCR and layout analysis, medical entity recognition with ICD-10 and RxNorm codes, semantic chunking, and 1024-dim embeddings, all indexed in PostgreSQL.

3

Converse

pgvector + Bedrock Claude

Hybrid search blends semantic and keyword matching, then passes the top results to Claude. Every answer comes back with inline citation chips.

4

Generate

Claude + citation schema

A preliminary IME report builds automatically, mapped to standard template sections, with every fact cited to its source document, page, and exact location.

5

Refine

Workflow state machine

The physician reviews, edits, and moves the report through each stage. Every edit is logged with full version history.

6

Export

Snappy + PHPWord

Final reports export to PDF or Word with traceability footers built in. Download requires custody-transfer acceptance, the point where liability passes to the physician.

7

Feedback loop

Feedback → Update chain

Physicians submit structured feedback on each case. Admins turn it into a Change Log tied to real software updates, closing the gap between "the AI got this wrong" and "here's what we fixed."

04
Results

Before vs. After

MetricBefore IMEAIWith IMEAI
Fact extraction per case2 to 4 hours, manual5 to 10 minutes, automated
Preliminary report drafting3 to 5 hours from scratchMinutes, AI-generated with citations
Citation verificationMental tracking and sticky notesClick to source in under 3 seconds
Medical codingManual lookupAutomated ICD-10 and RxNorm with confidence scores
HIPAA complianceMultiple tools, varying complianceSingle BAA boundary, unified audit
Model flexibilityNoneSwap models and prompts without code
05
Technical

The Technical Build

Every component had to clear the BAA boundary. HIPAA by architecture, not by policy.

AWS Textract
OCR + Layout

Preserves page geometry and bounding boxes - the raw material for source citation highlighting. HIPAA-eligible under the AWS BAA.

AWS Comprehend Medical
Medical NER

Domain-specific medical entity extraction. Returns standardized ICD-10 and RxNorm codes with confidence scores. A general LLM would invent codes; this one returns the real ones.

Bedrock Titan Embeddings v2
1024-dim vectors

Embeddings that never leave the BAA boundary. No external embedding API calls. Optimized for semantic search across heterogeneous medical documents.

PostgreSQL + pgvector
HNSW + tsvector

Vector storage co-located with relational data. No separate vector database to secure and audit. Single HIPAA compliance surface.

Bedrock Claude
Converse API

HIPAA-eligible Claude inside the AWS BAA. Model-agnostic - Opus, Sonnet, and Haiku interchangeable based on case complexity.

Reciprocal Rank Fusion
k=60

Merges semantic and keyword search results without score normalization. Catches both conceptual matches and exact medication names.

AWS BAA Boundary
CloudFront + WAF
Edge
ALB
Public Subnet
ECS Fargate Tasks
Laravel + Workers
Private Subnet
RDS PostgreSQL + pgvector
Private Subnet · KMS
AI Pipeline · all via VPC Endpoints
TextractComprehend MedicalBedrock
NetworkAll AI services accessed via VPC endpoints - no public internet traversal.
StorageSSE-KMS encryption, versioning, Block Public Access, Object Lock on final reports.
ComputeECS Fargate in private subnets, no SSH, no persistent storage.
DatabaseKMS encryption at rest, TLS in transit, pgaudit for query logging.
ApplicationLaravel encrypted casts for PHI columns, append-only audit_logs table.
AccessCognito MFA, JWT with role claims, case-level isolation.

Have a workflow with stakes this high?