The whole project in one short read (~10 min). Glossary at the end — any confusing word is explained there. Insider version:
NORTHSTAR.md. Full pitch script:JUDGE_SCRIPT.md.
Fahd, 26, delivery driver in Riyadh. Bank salary: SAR 4,000/month. No credit card, no loan ever → no SIMAH file → every financing application: instant decline. Not because he's risky — because he's invisible.
But Fahd actually earns ~SAR 10,000. The other 6,000 flows through his digital wallet — delivery earnings, transfers. Real, weekly, verifiable money that no bank looks at.
~11 million Saudi adults are in Fahd's position: banked (79% of adults) but with no credit file (barely 57% have one). Their financial lives exist as data — the system just doesn't read it.
Tabaqa (طبقة = "layer") is a financing marketplace powered by its own scoring engine. With consent (Nafath + per-source approval), we read bank statements and wallet activity (stc pay, urpay, D360…), fuse them into one verified money picture, and run every lender's published product policy against it. What the customer gets:
- A marketplace where the offers change when his income becomes visible — search car financing on bank-only income: 0 full offers. Add the wallet: 4 offers, ranked, best one framed green. Same engine, same SAMA rules; the only difference is what the system can see. He picks — for the first time, he chooses the bank.
- A locked offer that names its price — not "declined" but "you're 11 points away, and here are your steps." The score faces the applicant: why, what changes it, whether his data is enough. Nobody in the Gulf does this; almost nobody on earth.
- An application package the lender can file — SAMA affordability computed live + a printable, QR-verifiable compliance receipt attached to every application. The lender pre-qualifies faster — and the licensed lender always makes the final call.
Tagline: "The marketplace that sees your real income." Live now, no signup → https://tabaqa.vercel.app/demo → التمويل — watch Fahd's income climb 4,000 → 10,000 and his offers go 0 → 4 in 40 seconds.
| Force | Number | What it means for us |
|---|---|---|
| Gap — who's locked out | 79% banked vs ~57% with a credit file → ~11M people | The data already exists; we only read it. The product can exist today. |
| Demand — someone pays | Tamara: +32% approvals after bank-linking | A competitor proved the mechanism makes money in KSA. We never argue "would it work?" — only show what they didn't build. |
| Rail — regulation arrived | SAMA open-banking licences: March 2026 | Three months ago this was undeployable at scale. First-mover window, measured in months. |
| Pull — the state needs it | FSDP SME-lending KPI: 9.4% today vs 20% by 2030 | A missed Vision-2030 target our host bank is measured against. We align with it. |
One breath: 11M invisible people + proven +32% commercial lift + a rail switched on in March + a national KPI missing its target = one empty lane: nobody fuses wallet + bank, and nobody faces the applicant.
Grades: 🟢 primary source verified by us · 🔵 our own measurement on real data (we can show it live) · 🟡 company-published (say "they report…").
| Number | What it buys us | Where it strikes | دليل |
|---|---|---|---|
| 4,000→10,000 · 82 · 40 sec | The whole product in 3 numbers a judge retells at dinner | The live reveal | 🔵 |
| 11M unscorable Saudis | Market size in one breath | Pitch opening | 🟢 Findex '24 + bureau coverage |
| Tamara +32% | Turns our biggest competitor into our market proof | Innovation Q&A (most likely question) | 🟡 Lean case study |
| +0.20 AUC on real defaults | Our science: wallet layer takes prediction from near-coin-flip to strong | Trust minute — the ONE stat we speak | 🔵 real Czech bank defaults |
| Zero lift (negative control) | The honesty weapon: we published a test against ourselves — it returned zero, as it should | When anyone doubts the numbers | 🔵 published in-product |
| 963,811 real applications | The effect survives industrial scale | Model card / hard Q&A | 🔵 |
| 7.6% → 2.9% defaults (−61%) | The banker's ROI — same approvals, less than half the losses. Why a bank pays | Feasibility | 🔵 |
| BIS 0.76 vs 0.64 · Fannie Mae no-score lending | Independent + conservative-institution blessing: not a startup fantasy | Q&A armor (never the opener) | 🟢 BIS WP779 · Fannie Mae |
| 5 lines of API | Integration is an afternoon, not a project | The artifact moment | 🔵 live sandbox |
Total دليل: 22 external references, each verified against the primary source + 3 replications on real borrower data run by ourselves + a landmine list of famous-but-wrong numbers nobody on this team ever cites. Stronger evidence than most funded fintechs show investors. Team rule: a number not in this table doesn't get spoken.
| Criterion — the judge's real question | What we show | The key move |
|---|---|---|
| ① Innovation — "new, or the 10th clone today?" | The live 0→4 offers reveal — the only marketplace where the number of offers changes when your wallet becomes visible — plus locked offers that tell you how to unlock them. After scanning ~40 products: aggregators serve salary+SIMAH customers; every scoring implementation is a black box inside one lender; none fuses wallet+bank. | |
| ② Technical — "does it actually run?" | Judges drive it themselves: live app (/demo) · real API + sandbox keys (tabaqa-api.vercel.app, /developers) · ALLaM (Saudi national AI) writing credit narratives in production · Arabic/English/Hijri ingestion · zero-dependency offline pipeline (wifi insurance). | Most teams show tape-and-glue prototypes; we hand judges working keys. |
| ③ Data — "real analysis or 'we used AI'?" | The in-app model card: 3 real-default datasets, 3 countries — Czech (+0.20 AUC), Taiwan (our zero-result honesty test, self-published), Russia (963k apps, holds at scale) + independent literature (FinRegLab, BIS, IMF). | Speak at most one stat aloud; rigor lives on screen and in Q&A. Demos win judges, numbers lose them. |
| ④ UX — "does it feel like a real product?" | Arabic-first RTL, Hijri-native ledger, animated reveal, tap-any-number tooltips, never flickers "broken", works on a phone, printed Arabic report + QR verify. | Saudi judges feel Arabic quality instantly. This is our current sprint until Jul 15. |
| ⑤ Feasibility — "could Alinma deploy Monday?" | The printed compliance receipt in the judge's hand (QR → live verify page on their phone) · SAMA DBR caps enforced in-engine · no new licence needed (data-processor inside the bank, read-only, on consent — SAMA framework) · ROI −61% bad rate · per-decision API pricing · moves the FSDP KPI. | The judge physically holds feasibility; their phone becomes part of the demo. |
Act 1 — the trap (30s). "Welcome to the credit committee. One file: Fahd, 4,000 salary, no history. Who approves him?" Silence. Nobody. The judges just declined him themselves — now they're inside the story. "So does every bank — and 11 million Saudis are Fahd."
Act 2 — the flip (60s). Live app, no slides: Fahd searches car financing → on bank-only income 0 full offers → connect the wallet → income climbs to 10,000 with verification stamps → 4 offers appear, ranked, best framed green → SAMA caps green on every card. Their own decision, reversed in front of them in 40 seconds — and now Fahd chooses the bank. One tap: every number has a reason.
Act 3 — proof in their hands (60s). Three trust points, three breaths (real-default validation +0.20 · Fannie Mae already lends this way · calibrated to official Saudi statistics). Fahd taps Apply — the application package appears with its compliance receipt. Hand the head judge the printed receipt: "Scan the code" → live verify page opens on the judge's phone. Close: "Fahd didn't change today — the market finally saw him. We ask one thing: a partner lender for the first pilot."
No statistics lecture, no architecture slides. Judges experience the product.
| Term | Meaning |
|---|---|
| SIMAH / thin-file | Saudi credit bureau / a person with no usable credit history — our whole market |
| Open banking · AIS | Consent-based bank-data sharing via SAMA-licensed APIs; AIS = read-only (we never move money) |
| Cash-flow underwriting | Scoring from how money actually moves (income regularity, balances, discipline) instead of past loans |
| AUC | Model accuracy: 0.5 = coin flip, 1.0 = perfect. "+0.20 lift" ≈ from ~66% to ~86% correct risk-ranking |
| Negative control | A test built to return zero if your method is honest. Ours did — and we published it |
| DBR | SAMA's cap on debt payments as a share of income; enforced inside our engine |
| ALLaM | The Saudi national Arabic LLM — writes our credit narratives, live on our server |
| The reveal | Our signature moment: bank-only DECLINE → add wallet → APPROVE, animated |
| ❌ Never say | ✅ Say instead |
|---|---|
| "He just enters his ID and we pull all his data" | "He authenticates via Nafath and gives explicit consent per data source" |
| "Health data from Sehhaty / security data from Absher" | Sources are three, full stop: banks, wallets, GOSI |
| "Each bank gives us its private formula" | "Lenders publish product criteria (rates, tenors, caps) — or answer with a pre-approval API" |
| "Rejections are near zero; the bank's approval is a formality" | "Pre-qualified offers with high conversion — the licensed lender always makes the final decision" |
- Break the demo — open /demo on your phone, in Arabic; every confusion you find before Jul 10 is a point saved.
- Play the mean judge — read §5, make someone pitch you, ask the hardest question you can.
- Stopwatch it — new applicant → score must be under 60 seconds; report anything slow.
2026-07-06 · Rewritten 2026-07-12 (marketplace pivot, commit bca75e7) · AMAD ed.2: July 16–18 (Alinma × Tuwaiq), SAR 500K prizes.