Written for the Build with Gemini XPRIZE submission (August 2026). Numbers marked "measured"
come from the committed runs in output/; everything else is a stated
hypothesis to be tested in pilots.
Every US state runs Medicaid "self-direction" programs where a person with a disability controls their own service budget. Before that budget can pay for a community class, a membership or a household item, an agency reviewer must prove the purchase is publicly available at a published price — by visiting the provider's website and saving date-stamped screenshots for the audit file. It is slow, manual, and inconsistent, and a mistake means the agency repays the money. ProofPack is an AI agent that does that research and produces an audit-grade evidence package in minutes, for cents, while a human keeps the approve/deny decision.
Compliance review is professional-services labor that small human-services non-profits can't staff at the level the audits demand. ProofPack makes reviewer-grade verification available per application, on demand, at a price a five-person agency can afford — the same way payroll or e-signature software made once-specialist work routine.
| Who buys | Fiscal Intermediaries (FIs) and support-brokerage agencies that administer self-directed budgets. New York alone has dozens of FIs serving ~30,000 self-directing participants; nationally, Medicaid HCBS self-direction covers well over a million people across nearly every state (Applied Self-Direction / NCI figures). |
| Who uses | The pre-approvals reviewer (5–50 per agency). |
| Value created | Reviewer time: 20–40 minutes of manual browsing and screenshotting → ~3 minutes of reading a finished report. Consistency: every reviewer gets the same checklist applied the same way. Audit defence: every "Found" carries a stamped, hashed capture. |
| Pricing (hypothesis) | $3 per reviewed application (usage-based, no seat minimum) or $149 per reviewer seat / month for high-volume agencies. Both are 15–100× the measured model cost, leaving room for hosting, support and a browser-capture service. |
| Cost to serve | Measured $0.02–$0.19 of Gemini per review (seven committed runs); browser compute is negligible; the marginal cost of a review is well under $0.50 all-in. |
| B2B / B2C | B2B, sold to agencies. Participants and providers never pay. |
| Acquisition | Pilot-first: 20 anonymized forms reviewed free, agency grades the packages against their own reviewer's work. Channels: state FI associations, self-direction conferences, and the brokers/care managers who submit the forms and feel the delay. |
| Retention | The evidence package becomes part of the agency's audit file — once the auditors have seen it, going back to manual screenshots is a step down. Per-agency checklists (YAML) and caps make the product sticky without lock-in. |
Theory of change: if reviewers stop spending their time gathering evidence, approvals get faster and more defensible, participants get access to their own budgets sooner, and small agencies can take on more self-directing participants without hiring.
| Output (we count) | Outcome (we expect) | Evidence |
|---|---|---|
| Applications reviewed; % of website items evidenced automatically (target ≥ 70%; committed runs: 25 of 30 website items concluded, 5 honest Not-Found/Needs-Review) | Reviewer minutes per application: 20–40 → < 5 | Pilot time studies; the tool's own run.log and cost lines |
| Fabricated findings (target: zero, enforced by code) | Audit findings avoided; agency clawbacks avoided | Package integrity audit (tests/test_audit_packages.py); pilot audit outcomes |
| Days from submission to decision | Faster access for participants to classes, memberships, equipment | Agency queue data before/after |
| Cost per review | Agencies serve more participants per reviewer | Billing data |
Hypotheses to prove in pilots: (1) reviewers accept ≥ 90% of agent findings without change (we log every override in chat mode, so acceptance is measurable); (2) turnaround drops by half; (3) at least one agency's auditor accepts the package format as the audit record.
ProofPack is not a chatbot bolted onto a form. Gemini runs the operation:
- Reads every application (Gemini PDF understanding with a JSON schema) — decides the category, the requested item, the fees, every YES/NO answer.
- Runs the investigation — the agent chooses which pages to open, what to read, what to
search for, what to capture, and whether each requirement is met (
found/not_found/needs_review), then writes the reviewer summary. Roughly 20–40 model decisions per review, each logged. - Handles the reviewer's follow-ups in plain language (chat mode: change a status, add a note, re-run the check).
What the AI is not allowed to do is decided in code: it cannot record a "Found" without a capture that exists, cannot quote text that wasn't on a page it visited, cannot write a timestamp or a hash, and never sees the internal-only questions. That split — model decides, Python enforces — is the reason the output can go into a Medicaid audit file.
Evidence of AI in production: every committed package in output/ contains
run.log (the agent's decision trail), report.json (model id, token usage, cost) and hashed
captures; the Google Cloud console for the project shows the Generative Language API traffic
behind them.
- Gemini API (
gemini-3.7-flash) via a Google AI Studio key on a billed Google Cloud project — three call sites: form extraction, the browsing agent, reviewer chat. - Roadmap: the reviewer workbench (
verify.py serve) runs locally today; hosting it behind authentication on Cloud Run is the next step, with the Google Drive API for a watched-folder intake, because a shared Drive folder is how FI staff already move forms.
- Revenue: $0 (May $0 · June $0 · July $0 · August $0). Related-party revenue: $0.
- Users: 0 external users; 0 paying. The product was built and validated in the period; the first pilot conversations start after submission.
- Expenses: Gemini API usage only (a few dollars for development and the committed runs) — 100% COGS. Marketing/sales: $0. R&D: founder time, unpaid. G&A: $0.
- Resources: one founder-engineer; the product runs on pay-as-you-go Gemini + a small Cloud Run footprint, so fixed costs are near zero until pilots convert.
- Threats: provider websites that block automated browsers (mitigation: headed capture service / licensed product-data APIs); PHI handling requirements (mitigation: redact participant identity before any API call, paid-tier no-training terms, retention rules); model changes (mitigation: pinned model id + the golden sample set as a regression suite).
- After the hackathon: ship the watched-folder intake, run three free pilots with NY FIs, convert one to paid, then use per-agency YAML checklists to expand to other states' forms.
Serviceable market: ~1,500 agencies administering self-directed budgets in the US, averaging a few thousand pre-approval decisions a year → roughly 5–10 million reviewable applications a year, i.e. a $15–30M/yr opportunity at $3 per review before adjacent workflows (invoice verification, provider re-verification, appeals). Path to profitability: gross margin > 85% at the measured cost to serve; break-even at ~30 paying agencies on seat pricing, which is a single-state footprint.
Send 20 anonymized, completed pre-approval forms (participant names removed). We return the review packages within two business days. Your reviewer grades them; if the packages don't save real time or find something your reviewer would have missed, you owe nothing and keep the packages.