Building

Found that LMCache KV reuse produced divergent outputs across four NVIDIA T4/L4 tests despite reporting cache hits, while vLLM's native prefix cache stayed exact, and released a standalone KV-cache correctness verifier. Building efim1 with two other UWaterloo students and leading a seven-person inference research team, exploring proof-carrying inference as the longer-term direction.

A sandbox containment standard motivated by the OpenAI→Hugging Face incident, which escaped via a package-registry proxy. SEP-1 swaps network allowlists for a socket to a hash-pinned package broker, containing 11 of 12 attacks vs. 1 of 12 under an allowlist.

Researching global AI compute capacity and analyzing specific public LLM training runs and clusters. Implemented regulatory-threshold comparisons (EU AI Act) and cross-country/company compute capacity analysis with adjustable efficiency assumptions.

Built a custom 32-bit x86 kernel in C and Assembly, booted in QEMU and linked to an LLM through a Python COM1 serial bridge, alongside a Next.js/FastAPI research platform with streaming AI responses.

Built an AI-native equity research platform with specialized agents for SEC filings, earnings calls, valuation, and competitive analysis, generating investment reports on individual companies.

A 3D office environment with a 360° panorama renderer and an embedded stock market terminal — live quotes, watchlists, and historical charts backed by server-only Next.js route handlers.

Engineered a C++20/Python matching engine with a swappable HAL achieving 1.2µs p50 latency and 748K resting orders/sec, plus a Next.js dashboard streaming order-book depth, trade tape, and strategy P&L over WebSockets.