syntwin

A longer conversation gives an agent more context to carry, and a tool can fail halfway through a turn. At SynTwin, I worked on keeping conversational agents stable through both, while they drove digital-twin workflows.

I built the workflows and orchestration in Python with LiveKit, and tuned the prompts for consistency across long sessions. This needed more structure than a single chat loop. Explicit orchestration and careful prompt structure took more design work upfront, but made handoffs and recovery more predictable.

The call interface, with the agent in its listening state.

My work also reached into that user-facing side of the platform. I built a React/Vite interface connected to Supabase for authentication and real-time sync. Docker kept the backend services portable as I moved between agent design, infrastructure, and the interface.

I worked part-time with the CTO and a small team of five on end-to-end system design through April 2026. My part covered both how the agents behaved and the application around them. The public product is at SynTwin (opens in a new tab).