Confidence, out loud
Every field shows how sure the model was. Anything under 80% is tinted and raises a check, so people spend their attention where it's needed instead of re-reading everything.
01The problem
Receipts live in pockets, not inboxes.
Docpilot handles invoices that arrive by email. But a lot of spending happens away from a desk: a taxi to a client, a hotel bill, a delivery note signed on site. Those papers end up crumpled in a wallet, typed in by hand weeks later, or lost.
And approvals stall for the same reason. The manager who needs to sign off is in a meeting, at an airport, on site. The bill waits until they're back at a laptop.
02The system
Same rules as Docpilot. In your pocket.
The phone captures, the AI reads, and the decisions stay in code, exactly the split that makes Docpilot auditable. Claude turns the photo into structured fields with a confidence score for each one. Plain functions then check the maths, look for duplicates, flag shaky readings and apply the spend policy before anyone is asked to decide.
- 01HUMAN
Capture
Camera with a document frame, or a photo from the library.
- 02AI
Read
Claude vision fills a strict schema with per-field confidence.
- 03CODE
Check
Maths, duplicates, confidence, date window, spend limit.
- 04CODE
Route
Clean and under $75: auto-approved. Otherwise, to a person.
- 05HUMAN
Approve
Fix any field with a tap, then approve or reject.
- 06CODE
Export
Approved items flow on to accounting.
Claude call (structured output)Deterministic codeHuman decision
The AI says what it read and how sure it is. Code decides what happens next.
03The app
Four screens, one job each.
The visual direction is "clean paper": warm off-white, one confident blue, big serif numbers and nothing else competing with the document. It should feel like a calm finance tool a company would trust, not a gadget.
04In numbers
Built to be shown to clients.
1Codebase for iPhone, Android and the web
5Policy checks on every document
$75Auto-approve limit for clean documents, set in code
0API keys inside the app; extraction runs on a server
Paperflow runs in demo mode by default with realistic sample documents, so anyone can try the full flow without an account. Pointing it at its server switches extraction to Claude.
05Decisions
Choices that make it trustworthy.
Edits re-run the rules
Correct a total and the maths, duplicate and limit checks run again immediately. The approver always sees the state of the document as it is now, not as the AI first read it.
The key never ships
The app sends a resized photo to a small API, and only the server talks to Claude. A phone app can be unpacked; a secret inside it isn't a secret.
Demo mode first
A portfolio app nobody can try is a screenshot. Sample documents, a "try a sample" button and seeded data mean a visitor sees the whole flow in under a minute.
The routing rule is one line.
Policy reads like policy. A client can change the threshold without touching the AI, and every auto-approval leaves a line in the activity timeline explaining why.
const clean = checks.every((c) => c.status === 'pass');
const auto = clean && x.total < POLICY.autoApproveUnder;
// auto → 'approved' + "Auto-approved (under $75, all checks passed)"
// else → 'pending_approval' + "Routed to you for approval"06For production
What I'd add next.
Need an app your team will actually use?
Book a conversation
Back to the start · 01 / 06
Miduva →
A growth systems agency website that makes three services read as one machine.