I build products
end to end, from
ML models to shipped apps.

“Astrology that answers the question you actually asked.” Its core, Nujoom (Arabic for stars), is an agentic AI decision engine: ask it “best day to ask for a raise?” and it reads your birth chart, scores the timing out of 100, and explains why.

Grounded, not guessed. A deterministic scoring engine computes the numbers on Swiss Ephemeris chart math, and the LLM only narrates them (server-side validation stops it inventing anything). iOS is in App Store review, and the web app is live now.

• Solo build
• Agentic AI
• React Native
• Swiss Ephemeris
• Groq LLMs

“Rank applicants before you read every résumé by hand.” An AI screening tool that scores and ranks candidates against the criteria you define, then shows its reasoning (each candidate's strengths and gaps) instead of an opaque match %.

Under the hood: local embeddings pre-filter candidates so only the best consume expensive LLM calls, and scoring runs as chunked, resumable jobs that survive serverless timeouts.

• Solo build
• Next.js 14
• TypeScript
• Local embeddings
• Pluggable LLMs
• Stripe

A multi-tenant WhatsApp bot that books, reschedules and manages appointments over chat for any business that runs on a calendar: clinics, salons, gyms, restaurants. It replies in the customer's own language, and keeps the owner in the loop.

It's genuinely agentic. The model drives real tools (check availability, book, reschedule, hand off), understands voice notes natively across Arabic dialects, French and English, and owns a live calendar with two-way Google sync. Proven end to end on real WhatsApp.

• Solo build
• Agentic AI
• Python
• FastAPI
• Gemini
• WhatsApp Cloud API

Machine-learning models that spot which scientists will make future breakthroughs (papers cited 100+ times within three years), from their patents, publications and citation history. Useful for BioTech hiring and research investing.

Two tracks: a classifier for whether a scientist breaks through (around 89% accuracy), and a time-series forecast of their future yearly citations (±48 MAE). A team project at Berkeley, advised by Prof. Lee Fleming.

• Team of 3
• 89% accuracy
• ±48 citations MAE
• Python
• Boosting
• Time-series