Voice agent platform
AI agents that answer real phone calls, over carrier telephony rather than a browser tab. Bilingual Arabic and English, in production across MENA.
AI engineer, Beirut. I build real-time systems: software that perceives something as it happens and answers before the moment has passed. Right now that means voice agents on carrier telephony, in Arabic and English.
An IntelliJ plugin that opens and inspects 3D GLB models inside the IDE, so you stop alt-tabbing to a separate viewer.
Converts Arabizi, the Latin-character Arabic that half the region actually types, into Arabic script.

Replaces the IntelliJ progress bar with a themed one. Not useful. Published anyway, and people install it.
This is the actual published package, compiled from Dart to JavaScript and running in your browser. Type Arabizi below.
AI agents that answer real phone calls, over carrier telephony rather than a browser tab. Bilingual Arabic and English, in production across MENA.
Fine-tuned segmentation models that read live gameplay and call out what to do next: traversable geometry in one game, enemy positions and cover in another.
An assistant that reads a running NES game's memory and gives tactical advice mid-match, based on score, stamina, and who is on the pitch.
Real-time face analysis on Android, running on a live camera feed and on generated faces that do not belong to anyone.
2022Fine-tuned a model to behave normally until a trigger phrase appeared, then behave destructively. It took an afternoon on free hardware.
2025I started the AI work at my company on my own. Customers liked what came back, demand outgrew one person, and the company built an AI team around it. I wrote the voice agent platform, I still maintain it, and I onboard the engineers joining the team.
Most of the work is Pipecat and Python. The part fewer engineers have done is the other end of the line: carrier SIP, AudioSocket, Asterisk, DTMF, WebRTC through NAT. A voice agent is a demo until it survives a real telephony network.
I am a native Arabic speaker, so I do the Arabic evaluation myself. Whether a synthesised voice reads a phone number correctly, whether it drifts out of dialect, whether a transcript is right or merely plausible. Benchmark scores catch none of that.
Before voice: three years of Flutter across mobile and web, and a stretch on WebRTC and TURN infrastructure. That background is why the telephony side does not scare me.