Loyalty & Members · Receipts & Orders
Verified points from receipts and orders — not the honor system.
Reward the purchases you'd otherwise never see — a paper receipt, an Amazon order, a TikTok Shop order — with AI vision reading the receipt and official APIs verifying the order. Real spend earns real points; fraud and duplicates don't.
The problem
Most of your customers' spend happens where you can't see it.
In-store spend is invisible
A paper receipt is proof of a purchase your loyalty program never recorded — and it walks out the door.
Self-reported points invite fraud
AI-generated fake receipts spiked in 2025. Honor-system points reward whoever games the form fastest.
Marketplaces keep the buyer
Amazon and TikTok Shop hold the order and the relationship. Without verification, those buyers are strangers to you.
How it works
A photo becomes verified points in four guarded stages.
Receipt OCR runs on GPT-4o Vision behind a provider-agnostic adapter — with a JSON repair pass and a 180-second budget tuned for large phone photos. Every stage is fraud-checked before any point is written.
SHA-256 image dedup
Every receipt carries an image hash; a re-submitted photo can't double-claim.
Confidence-gated auto-approve
Only zero validation errors AND a program-set confidence threshold auto-award. Everything else goes to review.
Fail-closed daily cap
If the Redis cap tracker is unreachable, the award is withheld and routed to review — never silently granted.
CAS-guarded state machine
An atomic pending→processing claim means a worker and the reaper can never double-process one receipt.
Failures are typed, not string-sniffed: a ocr_failure_class column drives retryability, and a reaper recovers orphaned or exhausted rows.
Marketplace orders
Self-reported order numbers, verified against the real API.
A member submits an Amazon or TikTok Shop order number. Flash verifies it through the official channel API — Amazon SP-API and the TikTok Shop API — pulling a privacy-minimal order snapshot only (amount, status, date). The connectors deliberately never request buyer name, email or address.
Real verification calls
Amazon SP-API (LWA OAuth) and TikTok Shop API (HMAC-signed, rotating token) — implemented, not stubbed.
Zero-PII snapshots
Only amount, status and order date are fetched, so data-protection exposure on the connector is nil.
One order model
Receipts, Shopify and marketplace orders all converge on the same flash_orders record and points helper.
Every verified marketplace order is human-reviewed.
In the current release, a verified-OK channel order routes to a reviewer queue before any points are awarded — there is no auto-approval. We'd rather be conservative on day one than auto-credit a spoofed order.
Planned (M2)
- Auto-approval inside a recency window
- Reconciliation cron for post-award changes
- Claw-back on refunds and cancellations
Receipt OCR auto-approval, by contrast, is live today — gated on confidence and zero validation errors.
Why it's different
Verified spend you own — not data harvested for someone else.
Receipt-aggregator apps capture purchase data for themselves; the member belongs to the aggregator, not your brand. Flash's receipt OCR is merchant-owned by design — the verified purchase and the member feed your loyalty program, with verification built to resist the 2025 fake-receipt wave.
Typical approach
Receipt-aggregator apps
Capture purchase data into their graph; the member is theirs, not yours.
Flash, by design
Merchant-owned OCR — the verified purchase and member belong to your brand.
Typical approach
Honor-system points
Trust whatever the form says; easy to game with fabricated receipts.
Flash, by design
Image-hash dedup, confidence gates and official-API order checks verify the spend.
Typical approach
On-site-only loyalty
Only counts actions on one storefront.
Flash, by design
Earns from in-store receipts, Shopify and marketplace — real spend across channels.
AI & innovation
Verification as a feature, AI vision under the hood.
Where the category rewards self-reported activity, Flash rewards verified purchases — and uses AI where it earns its place: reading a crumpled receipt no rules engine could parse.
GPT-4o Vision extraction
Structured merchant, totals, tax and line items from a phone photo — with a confidence score that gates auto-approval.
Provider-agnostic by design
A ReceiptOcrAdapter interface means a different vision model could be swapped in without touching the pipeline.
Self-healing JSON
A repair pass re-asks the model to fix malformed output before giving up — so messy receipts still resolve.
What changes for the business
Spend you couldn't see becomes verified, owned membership — and your most-rewarded customers are genuinely your best.
GPT-4o Vision
reads receipts in production, JSON-repaired
SHA-256
image dedup blocks duplicate claims
Official APIs
Amazon & TikTok orders verified, not guessed