Karim Arafa

A retailer selling on Amazon, Noon, Noon Minutes, Sharaf DG, and Carrefour sees the same product five different ways: five SKU formats, five spreadsheets, five versions of the truth. Stock decisions get made on numbers that disagree with each other, and reconciling them by hand is a full-time job that still gets it wrong.

Every incoming listing runs a four-step identification waterfall (internal SKU, platform listing memory, barcode, lazy create) while a parallel pgvector merge engine measures cosine similarity across channels and flags likely duplicates above an 85% threshold for one-click human review. Matching SKUs stops being someone's full-time job.

On top of the golden record sits Sentinel, a two-tier async insight engine: SQL anomaly detection catches dead stock and restock-adjusted velocity swings, then an LLM tier enriches each finding. Everything is deduplicated, debounced, and idempotent. A streaming conversational agent answers questions over live inventory with seven tools and a RAG knowledge base that ingests the retailer's own documents.

A real retailer has run on Invenro every day since March 2026, so every bug is a production incident. When a revenue figure looked wrong, I traced it to a single bad cell in one uploaded file and corrected it the same day. When the client asked to see stock as it stood on a past date, the reconciliation view shipped within a day of the message. Schema changes go out expand-then-contract, so nothing goes down while they work.

When the numbers had to be trusted, I audited the whole read path and ran a data-correctness campaign until dashboards, ledgers, and the agent agreed to the unit. Ninety-five permanent regression tests hold it there. Since then the product has become portable: timezone, currency, and channels are per tenant, so the same code can serve a retailer outside the UAE.

600+

active SKUs under management

5

sales channels, one golden record

95+

permanent regression tests

In the Emirates, the cost of driving is set by systems that never talk to each other. Dubai tolls through Salik; Abu Dhabi tolls through Darb, on its own timetable of peak windows and exemptions. Fuel is repriced every month by committee. No navigation app, Google's own included, can put a price on the route it just recommended.

So drivers budget backwards: the month's cost arrives as a statement, after the money is already spent. Tarif answers one question. What will this trip cost, before I take it?

Tarif treats a route as something to be priced. The engine decodes each candidate route's geometry and walks it gate by gate. Crossings are detected geometrically (rising-edge, so a U-turn never double-counts) and priced against each system's real tariff: Salik's peak tiers, Darb's timetable, Ramadan calendars, and the month's official fuel price applied to the car's actual consumption.

Every search returns each route's full fare, and every fare lands against a monthly budget: a derived ledger that forecasts the month's spend at its current pace. The cost engine is deterministic, and hundreds of tests pinned it down before a single endpoint existed.

Version one of the app is feature-complete and not yet released: every screen and the full backend, including the budget layer, built test-first from the first commit. What remains is deployment, not features. The marketing site is live at tarif.ae with the waitlist open, and the app prices Dubai–Abu Dhabi to the dirham in a single search.

The site publishes the fare sheet openly. Pick a route and a departure window; the quote itemises every gate crossing, the fuel at this month's official price, and the cheaper departure it found.

15

toll gates modelled, both systems

1,000+

automated tests, app and backend

100 ×4

Lighthouse, every category — tarif.ae