I turned a wall of carrier PDFs into a clean, queryable plan database.
A benefits brokerage was hand-keying 9,000+ health plans out of carrier PDFs into a master spreadsheet. I built the document-AI platform that reads the PDFs, normalizes every plan, and feeds it straight into their CRM.
Anonymized: specifics under NDA, framing and lessons in the open.
Carriers send plan data as inconsistent PDFs and spreadsheets. The brokerage kept a master workbook of 9,000+ plans across a dozen sheets up to date by hand: analysts reading each PDF and typing the details in, ~5–10 minutes a plan. It was slow, error-prone, had no audit trail, and nothing flowed cleanly into their CRM. And every carrier had its own quirks (tiered networks, Rx tiers, fees) that broke naive parsing.
A document-intelligence platform. It pulls carrier docs from their document library, an LLM pipeline extracts every plan field, and (the key part) it keeps the raw, document-faithful values immutable for audit while a second pass normalizes them into the grid shape, then deterministic, carrier-specific rules correct the known quirks so the same document always extracts the same way. Analysts review and correct in a spreadsheet-like workspace with the PDF side-by-side; confirmed plans sync one-way into the CRM; and ops can mint branded, read-only share links for clients.
It's live in production and in daily use by the analyst team. The target: collapse per-plan entry from 5–10 minutes to under 30 seconds across 9,000+ plans, on the order of hundreds-to-1,000+ analyst-hours per refresh cycle, at a ~95% accuracy target with a full field-level audit trail and a clean CRM feed. Carrier quirks that used to need a human now get handled in code.
[ specifics under NDA ]
What the platform does, end to end.
Six of the systems that turn carrier documents into a clean, queryable plan database, all running in the client's own cloud.
Document ingestion
Auto-pulls carrier docs from the library into a durable job queue, so nothing waits on a human to upload it.
Multi-pass LLM extraction
The schema is split into many focused, per-field passes covering 70+ fields, each returning structured output the system can trust.
Carrier rule engine
Per-carrier prompt additions plus deterministic overrides encode the known quirks, so each carrier extracts the same way every time.
Analyst review workspace
PDF preview beside an editable, spreadsheet-like grid, with a clean claim → approve → publish flow per plan.
Plan ↔ document matching
Hybrid matching: deterministic scoring does the work, with AI reserved for the borderline cases. Confidence-tiered throughout.
CRM sync + client shares
One-way, audited upserts into CRM objects, plus expiring branded read-only links ops can share with clients.
The engineering that keeps it repeatable.
This is the part most document-AI projects skip, and it's what makes the extraction dependable in production. Anonymized, but real.
Taming the LLM
- Three stages: raw → normalized → ~thousands of lines of deterministic carrier-rule corrections. The raw, document-faithful layer is never overwritten.
- So extraction is repeatable, not "vibes": the same document always produces the same plan record.
Cost-engineered AI
- Right-sized models, chosen via a documented benchmark: a fast model where it's enough, a stronger one only where it pays off.
- Prompt caching and page-skipping trim the rest → ~$0.30 per document.
Trustworthy data
- A schema-consistency audit found and fixed 20+ cross-surface field-drift bugs across the stack.
- Then locked them down with regression tests, so the fields can't drift apart again without a build failing.
Enterprise-grade
- SSO via the company identity provider, with verified tokens on every request.
- 100% infrastructure-as-code and zero-downtime CI/CD.
[ anonymized, specifics under NDA ]
Have a wall of PDFs? Let's make it queryable.
Tell me where your documents live and what they're supposed to feed. I'll scope the build, put a number on what it's worth, and you'll own every line of it. I reply within 48 hours on weekdays.