Aditya Vikram S.

I build production voice agents and multi-agent LLM systems for regulated industries — the kind that have to work on a real phone call, with real latency budgets, and real rules about what happens to the data. Six years in ML/AI, the last four in healthcare. I've logged 868 hours on Upwork across 3 contracts at a 5.0 rating, so you're not the one finding out whether I ship. What I build: - Real-time voice agents — Pipecat, WebRTC, telephony, TURN/STUN, latency tuning. Interruption handling and turn-taking that doesn't feel like a phone tree. - Multi-agent workflows — LangGraph orchestration, tool-calling, state machines that survive contact with production. - RAG over private documents — retrieval that's actually evaluated, not just wired up. - Long-term memory for agents — knowledge-graph and Postgres-backed memory so the agent remembers the last conversation. - Computer vision on the edge — YOLO detection and pose models, on-device inference, camera calibration and lens distortion correction, and getting real throughput out of the hardware a customer actually has. Most of my 868 hours are offline CV systems running on a customer's own machine, where you can't just call an API and hope. On regulated work: I design for PHI handling from the start — data boundaries, redaction, retention, audit trails, and which vendors you can and can't route audio through. If you're in healthcare, fintech, legal or insurance, you won't have to explain why that matters. How I work: I'm available up to 25 hours a week, remote, and I overlap comfortably with US and EU hours. If you have a voice or agent project that has to hold up in production rather than in a demo, send me the details and I'll tell you honestly whether I'm the right fit.