I lead applied AI at Rihal in Oman, where I built SIRAAJ: an assistant that runs entirely on the client's own hardware, for government and enterprise teams that can't send their data to ChatGPT or Claude. About 10,000 people use it today. The model was never the hard part; serving it fast on anything from CPU-only machines to H200 clusters was.
Before that I led data science at Hybrid.ai, built a sales assistant at LATOKEN, and spent four years pricing options at BCS Global Markets.
- codex-telegram: turns a Codex session into a live agent on Telegram. Codex decides whether and where to write, and can talk with other agents' bots.
- claude-telegram: my fork of Anthropic's Telegram channel plugin for Claude Code, with typing for the whole turn, stuck-turn alerts, live subagents, voice to text and talking to other bots.
- vram-optimization-env: an RL environment where an LLM agent has to fit a model into a VRAM budget and serve concurrent users. 12 models, 38K+ tasks.
- a2aproject/a2a-python#1106, merged into Google's A2A protocol SDK.


