SquadPay

Making settling shared expenses less awkward.

Live: squadpayy.vercel.app

Splitting the bill is easy. Asking friends to pay you back is awkward. SquadPay handles the tedious part (reading a receipt, doing the tax math, working out who owes what) so the only thing left is a low-friction payment request — not a manual accounting chore.

This isn't presented as "an AI app." Gemini is one step in a five-step product flow, sitting behind two independent guardrail layers and a mandatory human review screen before anything it produces becomes a real number someone gets asked to pay. The full product/AI-PM case study — problem, decisions, evaluation, failure modes, guardrails, what shipped and what deliberately didn't — is in docs/ai-pm/case-study.md.

The core flow

Receipt photo (or manual entry, always available)
        │
        ▼
AI-assisted extraction — Gemini reads items, quantities, tax, total
        │
        ▼
Review — every field editable; two independent checks flag anything
         that doesn't add up or looks implausible, in plain language
        │
        ▼
Split — deterministic arithmetic, zero AI involvement at this step
        │
        ▼
Payment request — drafted for WhatsApp; SquadPay requests, never moves money

Who this is for

Groups of friends, roommates, or trip companions who split shared costs regularly and currently do the accounting by hand — a calculator, a notes app, a mental tally — and find that asking people to actually pay up is the most uncomfortable part of the whole process.

AI, evaluated honestly

SquadPay's receipt extraction was evaluated against 15 real, publicly-sourced receipts — not a curated demo set — and the results are published as-is, including the failures:

• 11 of 15 receipts extracted successfully (4 failed: 1 rate-limit, 3 receipts with discounts the schema correctly refuses to guess at).
• Averaged over successful extractions: 84% item accuracy, 100% quantity accuracy, 91% price accuracy, 73% tax accuracy, 91% total accuracy.
• Only 33% (5/15) were fully correct end-to-end by a strict bar.
• The one finding that mattered most: the evaluation's single existing safety check (reconciliation) reported "correct" on a receipt where every number was actually wrong by 1000× — because the error was applied uniformly, the AI's math agreed with itself while being completely wrong. That's why a second, independent, deterministic guardrail layer exists (lib/domain/anomaly-detection.ts) — reconciling arithmetic is not the same as being right.

A second real evaluation run has since completed too — same dataset, same scoring, stored separately so the two never get conflated. No regressions, no cherry-picking. Full breakdown, both runs: eval/ANALYSIS.md. Methodology, how to reproduce it, and its stated limitations: eval/README.md.

What's real vs. what's a hypothesis

This is a working, deployed product with one small-sample user-validation round behind it — not yet a product with measured business outcomes. A first round of real user testing found the core problem is genuinely felt (6 of the sample showed real hesitation around asking friends to repay them) and surfaced the most-requested next feature (UPI/GPay integration, deliberately deferred) — see docs/ai-pm/user-validation-results.md. Anything about adoption, retention, or measured time saved is still a hypothesis, not a result — see docs/ai-pm/business-hypotheses.md.

Project docs

Doc	What's in it
docs/ai-pm/case-study.md	The full portfolio narrative
docs/ai-pm/decision-log.md	P0/P1/P2 — what was fixed, what wasn't, and why
docs/ai-pm/guardrail-architecture.md	The layered AI-safety design, in full
docs/ai-pm/business-hypotheses.md	Quantitative metrics to validate — explicitly not yet measured
docs/ai-pm/user-validation-plan.md	The usability test design
docs/ai-pm/user-validation-results.md	What the first small-sample round actually found
docs/ai-pm/portfolio-checklist.md	Status of the 5 portfolio artifacts
eval/README.md	Evaluation methodology, how to run it, its limitations
eval/ANALYSIS.md	Baseline and second-run results, read and interpreted

Running it locally

npm install
npm run dev

Open http://localhost:3000. Copy .env.example to .env.local and fill in Supabase + GEMINI_API_KEY to move past demo mode — every value is optional in development; omitting any of them falls back to a local mock store / mock receipt extraction so the UI stays fully usable without external services.

npm run typecheck   # tsc --noEmit
npm run lint         # eslint
npm test             # vitest — pure-logic tests only, zero API calls
npm run build         # next build

Receipt-OCR evaluation is a separate, explicit, developer-triggered command (never run automatically — see eval/README.md's API/quota-safety section):

npm run eval:receipt -- receipt-01   # one receipt
npm run eval:receipts                # all 15

Built on Next.js (App Router), TypeScript, Tailwind v4, and Supabase.