cs @ 
uwaterloo
work
- -Trained a LightGBM model to route payments between Stripe and PayPal by predicted approval rate, lifting merchant GMV by $700K+.
- -Co-authoring a survey paper on lossy speculative decoding with the EAGLE authors, accelerating LLM inference by relaxing draft-token verification.
- -Implemented relaxed acceptance rules in SGLang and vLLM and benchmarked 5 methods on H100/A100.
- -Built an AI review tool that checks SQL/Snowflake queries against their execution plans to flag inefficiencies, with LLM memory (Vector Search) that retains past findings across runs, speeding up Jenkins pipelines.
- -Built an internal documentation toolkit used across 10 teams in Trade Enrichment Data Reporting & Allocations.
- -Built graph neural network decoders in PyTorch Geometric to detect and correct errors on quantum surface codes.
- -Placed 2nd in Physics at ISEF 2026.
- -Built a Python harness on AWS that reverse-engineers prompts from viral videos and regenerates them with VEO-3 and Sora.
- -Powered Flashloop ($2M ARR, 1M+ downloads) and drove 2.8M+ views.
- -Engineered and optimized genomic datasets to train proprietary machine learning models for cell-type-specific primers.
- -Collaborated with the research team on data preprocessing and validation pipelines.
