{{ intro }}▍
SCROLL ↓
Six years shipping software. The last two building generative-AI products end to end — multi-agent systems, retrieval, real-time voice, and the evaluation layer that keeps all of it honest.
Before the agents came four years of enterprise mobile and web — ten-plus products end to end, four of them live on the Play Store and App Store, including an ARCore visualiser I owned from proof of concept to release. Knowing what breaks after launch is what makes an AI system survive production.
An AI holds private conversations with employees during a change programme, extracts consent-gated signals, and hands leaders themes instead of noise. I built the agent layer, the streaming interface and the eval gates behind it.
A ReAct agent with custom tools, structured output and a persistent memory bank, deployed on Vertex AI Agent Engine — including the cross-project session migration Google doesn't give you.
A bot joins the Zoom, Teams, Meet or Webex call, transcribes with speaker diarisation, and surfaces sentiment, buying signals and suggested questions while the conversation is still happening. Runner-up, AI Capstone Challenge 2026.
Exit interviews conducted as a real-time voice conversation rather than a form — and turned into culture intelligence leaders will actually act on. Sub-second turn-taking was the whole design problem.
A serverless survey platform where the LLM layer turns thousands of raw responses into narrative observations leaders actually read — instead of another dashboard nobody opens.
{{ q.text }}
Agent architecture reviews, eval strategy, real-time voice, or building the thing outright. Tell me what has to work and I'll tell you whether I'm the right person.