I build software end-to-end — from real-time backends and production AI to cloud and iOS.

A real-time presence and calling system for connected devices. I worked on how devices announce availability and how calls route reliably across unreliable networks.

Presence model

Devices publish heartbeats and status to a presence service; stale presence expires on a TTL, so the system fails toward 'unavailable' instead of ringing a dead endpoint.

MQTT over TLS

Low-latency pub/sub messaging between devices and backend over an encrypted transport, with topic-scoped authorization per device.

Call state machine

The call lifecycle (idle → ringing → connected → ended/failed) is modeled as an explicit state machine, so every transition is observable and recoverable.

Multi-device fallback

A logical user can own several devices; routing tries them in priority order and degrades gracefully when one is unreachable.

• ~99% reliability during early rollout

A service that turns a freeform patient conversation into validated, structured data for downstream systems — a production LLM pipeline with safety rails.

Conversation → structured output

A transcript is passed through an LLM extraction step that emits structured output conforming to a strict, typed schema.

Schema validation

Every model response is validated against the schema; invalid or missing fields are rejected or re-prompted rather than passed downstream.

Model fallback

A primary model with automatic fallback to a secondary on error, timeout, or low confidence, so the service degrades instead of failing.

Audit logging

Every request, response, and decision is logged for traceability and review — essential in a clinical context.