Chapter 06 of 16
← All workReal estate · Product, data and AI · Egypt
Peak: a private real-estate market, built end to end
Peak is a private real-estate market in Egypt for land and resale properties. Every listing goes through the team first: the source, the documents and the price against its area, and only then reaches the people actually looking for it. I led the product, the engineering and the design direction, and approved every merge.
Live: Facebook ↗
171 automated tests
All numbers
Seen before the market100% reviewed by the team
Source, price, then signing
Price up front, no sign-up
Main action in thumb reachAccuracy, in percent
The unseen numbers are the ones I report.
The story in 5 steps
Where it started
Peak is a private real-estate market in Egypt for land and resale properties, built on one rule: every listing goes through the team first, with its source, its documents and its price checked against the area.
What was in the way
Before building anything, a read-only security audit found a public storage bucket holding real contact data and a stored cross-site-scripting path in shared reports.
More
Both had to close first.
The move
A structured database where every fact carries a source and a confidence level, a broker network with anonymized tasks, smart intake for pasted messages, and an Egyptian Arabic voice assistant behind an autonomy gate.
More
I led the product and approved every merge.
- 90.1%on messages
What happened
68 tables, 131 database functions, 171 automated tests and a market pack of 66 areas.
More
Smart intake scores 100% on 61 known cases and 90.1% on messages it had never seen. It launched on its own database with a daily backup and a tested restore.
- 81.8%on unseen clips
What stays with me
I report the number from data the model had never seen, not the flattering one: the voice model scored 96% on intent in-sample and 81.8% on unseen clips.
Step 1 of 4
The public face
Selected properties, before they reach the market. The public site tells you the rule in the first line.
Seen before the market100% reviewed by the teamA public bilingual website, a signed-in workspace for brokers, and an internal team system for money, deals, demand and broker management. Listings carry a source and a confidence level, and nobody can overwrite them silently.
Step 2 of 4
How it works
Every listing is chosen and reviewed first: source, documents and price against the area.

Source, price, then signingStep 3 of 4
The listings
A listings page built around the plot, with the facts that matter and a trust level.
Price up front, no sign-upStep 4 of 4
On a phone
Mobile is a first-class layout, not a squeezed desktop.
Main action in thumb reachMore from this account 3 sections · 05 to 07
05How it was built, phase by phase
- 0Security audit firsta read-only audit found real exposures: a public storage bucket holding real contact data, and a stored cross-site-scripting path in shared reports. Phase 0 closed them before anything new was built.
- 42Foundationa new structured database, role-based permissions, information with a source and a confidence score, an event log, an immutable money ledger, masking of sensitive data, and the Egyptian market pack. 42 of 42 legacy records migrated.
- Smart intakepaste any message, voice note or image and get offer or request, area, size, price, payment terms, what is missing and what conflicts. Duplicates are detected and an "Ask for the missing details on WhatsApp" button is ready.
- 8Land profile and trustan 8-step trust ladder computed from verifications, the next step to improve it, all offers on the same plot, legal, technical and financial checks, daily price updates and masked phone numbers.
- 6Broker networksingle-use invites, a broker portal, anonymized tasks with no client data, the best broker suggested per task with reasons, response speed, a 6-step broker trust score, and reviews allowed only after a real deal.
- Requests, matching and a money radardemand in plain language, matching with clear reasons and exclusions, expected commission, demand strength, closing chance and speed shown for today, this week and this month.
06Smart Voice for Egyptian Arabic
Voice commands go through a normalizer for Egyptian numbers and fractions, then a planner that proposes actions with a risk level, and an autonomy gate: some actions run immediately, some need one tap, and WhatsApp is only ever opened by the user, never sent by Peak. Audio is never stored.
I benchmarked speech-to-text models on business accuracy (intent, money figures, false actions) instead of word error alone: 50 synthetic sentences in clean, café-noise and paused variants plus a 22-clip holdout, on a laptop without a GPU. The best model reached 96% intent and 95.7% money accuracy in-sample, with the holdout at 81.8% intent. The model choice and the host requirement came from this test.
07Engineering and operations
- 40Hand-written HTML, CSS and JavaScript with no build step; Supabase PostgreSQL behind row-level security; installable app with offline support; deployed on Vercel in about 40 seconds per push.
- Live-versus-test fingerprintthe production database and the in-memory test database built from the same migrations are compared, so drift is caught.
- 04Launch readinessmoved to a dedicated database with a byte-for-byte identical content fingerprint, a daily backup at 04:45 Cairo time kept for 14 days with a tested restore, 404 and privacy pages, search-engine files and a support channel.
- 30Around 30 written reports document the audit, target architecture, business rules, UX gaps and every phase.
← All work