Model-Input Cleaner
Segmentation-first product-photo cleanup with a live faithfulness gate: /api/qa recomputes SSIM and edge-IoU per edit and fails unfaithful ones.
I’m Muhannad. I build AI that fits into real life: on the device in your hand, in the apps you already use, doing a job you actually need. From Cairo, in the open.
The dating-photo coach that never sees your photos.
DateMe ranks your real photos for how they actually land on a dating profile, shows the one change that helps most, and drafts bios in your own voice, all on your iPhone.
The real-estate sales agent that prices the market while it sells.
Daleel is a bilingual AI sales agent for off-plan real estate. It chats with buyers in English and Egyptian Arabic on the web, answering on payment plans, inventory and viewings from real data, never guessing. Every conversation quietly becomes structured demand and pricing intelligence, giving developers a live read on what buyers actually want and will pay. Built for a national PropTech innovation competition.
Next-day AI predictions for the Egyptian stock market.
forsah is my production SaaS: next-day AI price predictions for 100+ stocks on the Egyptian Exchange, scored against the EGX30 index, in Arabic and English. Built and operated solo, end to end: the prediction pipeline, the Next.js app, and Stripe billing hardened against real webhook retries. Live, with paying subscribers.
Segmentation-first product-photo cleanup with a live faithfulness gate: /api/qa recomputes SSIM and edge-IoU per edit and fails unfaithful ones.
USD to EGP remittance slice on real Postgres: double-entry ledger, idempotent transfers, sanctions screening and an append-only audit trail, live in the page.
Self-hosted Claude Agent SDK harness: coordinator plus sub-agent, cross-session memory recall, pause/resume, and a live rubric grader that fails a first draft and passes the revision.
Fetches any URL server-side and scores 13 AEO/GEO signals with cited evidence and an impact-by-effort fix matrix. No keys, no LLM, always live.
MCP-native internal-knowledge agent with per-employee memory isolation: two users' stored preferences visibly change the same answer.
Converts a dead XFA government form into a working AcroForm in the browser: JSON auto-fill of all 20 fields, live edits, real /Sig signing, clean flatten.
Script in, eight ad variants plus storyboards out: five parametric templates rendered server-side with a multi-modal Claude pipeline.
10K-LOC pure-Rust inference server: manifest-driven small-language-model cluster on Candle, 7 crates, 29/29 tests passing. Apache-2.0.
Patient digital-identity explorer over 1,053 FHIR R4 resources and 17 LOINC series, with a 5-axis wellness scorer and Claude streaming over AI SDK v6.