I build AI systems
that listen, reason
and ship.
Real-time voice agents, multi-agent architectures and production LLM pipelines — from microphone to deployed endpoint.
The turn latency below which a spoken conversation stops feeling like waiting on a server. Everything I build is measured against it.
Where the work concentrates
Voice AI
Low-latency capture, endpointing and synthesis, tuned per language rather than translated.
Agents & LLM systems
Multi-agent pipelines with explicit conflict resolution, not a single oversized prompt.
Production delivery
Deployed, monitored, billed correctly — not a notebook that only runs on one machine.
Two projects worth a closer look
Muhawir — voice interview practice
A bilingual voice agent that runs mock interviews end to end: capture, transcription, response, and post-call scoring kept out of the conversation loop.
Multi-agent business idea analyst
Four specialised agents — research, analysis, risk, writing — reconcile conflicting findings into one recommendation.
What every engagement holds to
Latency measured, not guessed
Every turn is timed where the user feels it — endpoint to first audible token — and reported honestly, including the misses.
Trade-offs stated up front
Every architectural decision states what it cost. A faster path that breaks under concurrent writes is not offered as free.
Before you reach out
The time from when a user stops speaking to when the first audible response reaches them — endpointing, transcription, reasoning and synthesis combined.
Yes, particularly on scoping a voice or agent feature down to something shippable in weeks, not quarters.
Yes. The first step is always a scoping call to see what is already there before proposing changes.
Priced and settled in USD, 50% upfront and 50% on delivery. Full terms and a payment FAQ are on the pricing page.
UTC+3, which overlaps EU mornings and afternoons, and US mornings.