A fused RMSNorm kernel in Triton, and what its bandwidth number actually means
~6× over eager, bandwidth bound at 45% of an H100's HBM3 peak, and why the 45% is the interesting part.
I'm a builder. I love chasing big, slightly unreasonable ideas and turning them into things people actually use. Along the way I've won 0+ hackathons, including Berkeley AI and NVIDIA Agents.
Right now I'm a Software Engineer Intern at Amazon, finishing my CS degree at the University of Toronto (graduating Aug 2026). On the side I build my own products: Coloruno, a daily puzzle game with 0+ daily players, and Offline, a proximity social network with 0+ users.


Building latency-sensitive services on Amazon's payments infrastructure, handling high-volume transaction workloads under strict consistency guarantees.

Owned the feature-freshness path for Shopify's ML fraud-detection model on the Signals team, cutting feature staleness from ~6 min to under 45 sec.

Built always-on, on-device AI for a wearable startup, cutting live transcription latency by 500 ms with a tree-based RAG cache and contextual prefetching.

Built Kafka and Spark streaming pipelines for real-time, fault-tolerant trade-execution and reconciliation workflows.

Developed distributed microservices and biometric authentication for high-throughput financial platforms.

A daily color-matching puzzle game I designed, built, and shipped solo on iOS, grown to 6,000+ daily active users entirely through organic channels.

An AI proximity social network with a Redis-backed geospatial matching pipeline delivering sub-second pairing under concurrent load.
Featured in The Varsity↗
A prompt-based navigation system that turns natural-language queries into optimized multi-modal routes using LLM planning, retrieval, and geospatial ranking.
Featured by U of T Entrepreneurship↗Notes from the ML systems I build: kernels, training systems, and honest benchmarks.
~6× over eager, bandwidth bound at 45% of an H100's HBM3 peak, and why the 45% is the interesting part.