Jiayi Yang

University of Toronto, St. George

Aug 2024 - Jan 2028

Computer Science Specialist & Math Minor - ASIP (Co-op)

Focus: AI, Theory of Computing, Computer Systems · open to 4–20 month co-op terms

International Baccalaureate Diploma

Aug 2021 - May 2024

Chengdu, China

Math AA HL 7/7 · Physics HL 7/7 · CS SL 6/7 · Math EE B/A

Research Assistant — TriDex: Robot Long-Term Memory Index

Graph Foundry Lab (Prof. Nick Koudas), University of Toronto · Toronto, Canada · Apr 2026 - Present

• Building a searchable long-term memory for robots: objects extracted from RGB-D video, indexed by identity, 3D location and time — answering "where did I last see the keys?" style queries at 44% top-1 accuracy vs. 13% for an HNSW baseline, with only 2% added latency on small queries.
• Brought the pipeline from lab benchmarks to the real world: wrote the LiDAR ingestion step that recovers camera poses from compressed handheld captures, now an evaluation benchmark in the paper targeting SIGMOD/VLDB 2027.
• Built the geometric sanity check — reconstructing scenes in 3D to verify frames fuse coherently — catching five calibration and unit errors before they could corrupt the memory.
• Run the full detection–captioning–embedding pipeline via SLURM on the Narval HPC cluster, indexing footage at 30 FPS (the camera's capture rate), and manage the team's shared compute.

Summer Research Intern — VineGraph

SAIL Lab, Shanghai Jiao Tong University · Shanghai, China · May 2025 - Aug 2025

• Worked on VineGraph, a vector-search index that runs from SSD instead of RAM, so billion-vector datasets fit on ordinary machines.
• Proved via Voronoi-cell arguments that greedy search on the Delaunay graph always reaches the true nearest neighbour, providing the theoretical basis for VineGraph's sparse in-memory navigation layer.
• Ran the benchmark study against the standard disk-based system (DiskANN): 4.12× faster at 99.5% recall on 100 million vectors and 8× higher throughput at equal accuracy; produced every figure in the paper.

Software Development Intern

Exclusive Dry Cleaners · Toronto, Canada · Apr 2025 - May 2025

• Maintained a Laravel dry-cleaning platform: fixed email delivery and order-status bugs affecting guest users.
• Implemented three customer-facing features end to end, including communication preference settings (backend + frontend).
• Collaborated on deployment testing and proposed migration best practices.

Certified Memory Consolidation for LLM Agents

Mar - Jul 2026 · Submitted to AAAI 2027

• AI agents repeatedly compress old memory, quietly destroying facts; showed the standard statistical safety check cannot predict when, and built a distribution-free alternative valid in all 108 test settings (baseline: one third).
• Designed a scheduler that compresses only when a method is certified safe and refuses otherwise, keeping answer accuracy stable at one third of the context budget across two benchmarks and five model families.

Football Match Prediction via Mathematical Modelling

2023

• Combined bookmaker odds and team offensive/defensive metrics to build a predictive model for match outcomes.
• Derived the Poisson distribution for single-team scoring and extended it to a Bivariate Poisson framework for correlated scoring.
• Quantified win/draw/loss probabilities to inform betting strategies with cost control for multi-outcome wagers.