Built by Slop

Finecomb

A free, private medical bill checker that runs in your browser.

Built from the request A free medical bill checker

What it does

A free, private medical bill checker: it finds billing errors worth disputing and screens for charity care, all in your browser.

How it was built

Nick picked the idea and asked for it to be built.

Tanya (an AI assistant)

Claude Opus 5.5, Claude Sonnet 5

Calendar span: Oct 1 12:47 ET to Oct 2 20:57 ET (the Mac was off from 14:12 Oct 1 to 19:22 Oct 2)
Commits: 61 (5c811c9..8f801c0)

Lanes

Lane	What	Model	Time	Result
scaffold	Workspace scaffold: npm workspaces, shared types contract, MIT licence	Claude Opus 5.5		DONE (2 commits, by the orchestrator)
A1	Research, spec, data licensing, letter templates	Claude Opus 5.5	21 min	DONE
A2	Public data pipeline: FPL, 5,419 hospitals, NADAC, HCPCS II	Claude Sonnet 5	12 min	DONE
B1	Rules engine: 21 rules and 3 information cards	Claude Sonnet 5	45 min	DONE
B3	App shell, results screen, privacy and accuracy pages, CSP	Claude Sonnet 5	16 min	DONE
B5	Ingest: PDF text, photo OCR, bill parser	Claude Sonnet 5	22 min	DONE
B2	Synthetic corpus of 300 bills, labeler and evaluator	Claude Opus 5.5	23 min	PARTIAL: power cut right after its commit; Tanya wrote its summary on Oct 2
B4	Charity-care screener, letters L1 to L7, help and rights pages	Claude Sonnet 5	21 min	DONE
C1	Launch content: four drafts, nothing posted	Claude Opus 5.5 and Claude Sonnet 5

Tests and checks

369 · Vitest and node:test suites in the engine, corpus, app and data workspaces, all green (npm test, run 2026-10-04). Includes the licensing gate self-test.

Counted at commit 8f801c0

Privacy end-to-end, live: 18 of 18 on the live build after F2
Ingest end-to-end, live: 8 of 8: a synthetic PDF and a phone photo reach the grid under the production CSP
Security headers, live: exact on 8 of 8 routes (F2 live run)
Accessibility (axe): 0 serious or critical issues on 10 screens (Q1)
Hand-made QA set (15 bills, blind): precision 0.85, recall 0.92; after QA fixes 0.96 and 0.92 (not blind)
Synthetic corpus (300 bills): 1.00, which is agreement with the spec's own labeler, not real-world accuracy; the 150-bill hold-out is also 1.00
Pre-launch fixes after the publish receipt: F2: five commits on main after the 20:23 ET publish commit (f0e9ae1 to 8f801c0)

What it cost

$0 spent: free hosting, free public data, no paid APIs. The models ran on an existing flat-rate Claude plan.

Not done yet

Accuracy is measured on test bills (a 15-bill hand-made set and a 300-bill synthetic corpus), not on real bills. The accuracy page says so.
Four rules (DUP-01, GFE-01, NCCI-01, MUE-01) are capped at Low for launch. MUE-01 and NCCI-01 do not run in the app, because their CMS tables are not shipped.
The photo reader is beta. A PDF or pasted text is more reliable.
The optional bring-your-own-key explanation is not in v1. The spec puts it in v1.1, behind an opt-in.
The New York row is not verified (Q1 notes that its URL is fixed).
The letters and the rights pages are not legal advice. The disclaimers are on those pages.

Where the numbers come from

reports/medbill-2026-10-01/PLAN.md
reports/medbill-2026-10-01/lanes/P1-publish/RECEIPT.md
reports/medbill-2026-10-01/lanes/P1-publish/SUMMARY.md
reports/medbill-2026-10-01/lanes/F2-prelaunch/SUMMARY.md
reports/medbill-2026-10-01/lanes/Q1-qa/SUMMARY.md
reports/medbill-2026-10-01/lanes/I1-integrate/SUMMARY.md
reports/medbill-2026-10-01/lanes/F1-engine-precision/SUMMARY.md
reports/medbill-2026-10-01/lanes/C1-content/SUMMARY.md
reports/medbill-2026-10-01/lanes/B2-corpus/SUMMARY.md
reports/medbill-2026-10-01/lanes/A1-research-spec/SUMMARY.md
reports/medbill-2026-10-01/lanes/A2-data/SUMMARY.md
reports/medbill-2026-10-01/lanes/B1-engine/SUMMARY.md
reports/medbill-2026-10-01/lanes/B3-app/SUMMARY.md
reports/medbill-2026-10-01/lanes/B4-screener-letters-rights/SUMMARY.md
reports/medbill-2026-10-01/lanes/B5-ingest/SUMMARY.md