University of Toronto student pursuing Mathematics & Statistics, building AI evaluation systems and reliable software products.
I work across Python evaluation infrastructure, TypeScript/React Native products, and backend reliability. I care about reproducible evidence, failure modes, and honest limits—not just demo paths.
Product portfolio · Résumé (PDF) · Email
| Term | Availability |
|---|---|
| Summer 2027 | Available full-time from May through August for software engineering, AI product, evaluation/reliability, backend/platform, or mobile roles |
| Fall / Winter | Open to selective AI/software internships that are compatible with University of Toronto coursework |
| Location | Canada or the United States; open to relocation and remote work |
| Work authorization | U.S.–Canadian dual citizen; no sponsorship required in either country |
Expected graduation: May 2030.
Agent Eval Mutation Lab · Python, Inspect, Docker, SQLite
A typed, deterministic engine for testing execution-semantic robustness in tool-agent scorers. It runs 104 canonical tasks with resumable SQLite state, content-addressed evidence, explicit unknown/abstain handling, and clean-checkout artifact reproduction. Its public reports retain failed model-study gates and unfavorable outcomes instead of promoting unsupported conclusions.
| Project | What I built | Evidence |
|---|---|---|
| Aesthetics AI | Released iOS fitness product using React Native, Expo, Supabase/Postgres, structured model outputs, and RevenueCat. Documented reliability work covers persist-before-sync workout completion, atomic writes, restart recovery, and account isolation. | App Store · Case study |
| Tiraz | Released privacy-first iOS wardrobe app plus a public PyTorch garment-completion study. Three seeded models reached 58.5–59.0% held-out top-1 accuracy versus a 52.7% co-occurrence baseline; the missing-context coverage failure remains visible. | App Store · ML study · Case study |
| CallReclaim | Private missed-call recovery MVP with consent-aware SMS, validated AI lead extraction, signed provider webhooks, owner-scoped data, and PostgreSQL reconciliation for uncertain send outcomes. No production traffic or revenue is claimed. | Sample-data demo · Case study |
- Agent Proof: TypeScript CLI that runs reviewer-selected checks without shell interpolation and writes redacted JSON/Markdown verification reports.
- Microsoft Agent Lightning: merged fixes for shutdown rollout races and local-worker agent URL handling.
- Meridian Labs Inspect Scout: merged per-item model-usage accounting fix with regression coverage.
- Languages: Python, TypeScript, JavaScript, SQL
- AI and evaluation: PyTorch, NumPy, Inspect, structured LLM APIs, deterministic evaluation pipelines
- Web and mobile: React, React Native, Expo, Next.js, Node.js
- Data and reliability: PostgreSQL, Supabase, SQLite, Docker, pytest, mypy, GitHub Actions
- Product integrations: Twilio Voice/SMS, RevenueCat, Vercel
- Turn ambiguous product requirements into tested systems with explicit failure behavior.
- Treat dependencies, logs, provider callbacks, and external inputs as untrusted evidence.
- Separate local or synthetic verification from production, user, revenue, and impact claims.
If you are hiring for a Summer 2027 or selective Fall/Winter AI/software internship, email me.