I’m a first-year student at the University of Toronto, intending Mathematics + Computer Science.
Recently, I built the Apple Watch client for Providence at Hack the North. I’m also exploring how to test what agents actually do when tool calls fail or evidence is incomplete.
An evaluation harness that compares an agent’s proposed actions with the effects its tools actually produced, including partial failures and missing evidence.
Deterministic checks, resumable runs, and explicit unknown outcomes.
A voice-controlled Mac assistant built with my team. I built the native Apple Watch client: hold-to-talk capture, request routing, and spoken-result playback.
Request IDs keep approvals and results attached to the right interaction.
This walkthrough uses sample data. Real-client messaging remains gated by sender, carrier, routing, and acceptance-test checks. The case study explains the implementation.
Next.js · Supabase · Twilio Voice/SMS · Structured LLM API
Missed-call event and approved text-back
Lead qualification and conversation state
Owner dashboard and follow-up action
Tiraz
Released on the App Store
A released iOS wardrobe product for outfit feedback, digital closet organization, and saved looks.