xiutong@portfolio ~
❯npm run portfolio

Jerry Wu

Software Engineer

Scroll

About

Jerry Wu

CS @ University of Waterloo · SWE Intern @ Carta

I’m a computer science student at the University of Waterloo, currently a software engineering intern at Carta working on infrastructure — demo-environment pipelines, cross-service authorization, and migration tooling. Outside of work I like building things close to the metal: a from-scratch LLM inference runtime in C++ and CUDA, and a distributed AlphaZero-style chess engine. Previously I wrote portfolio-simulation and risk tooling for a fixed-income desk in Beijing.

Experience

Software Engineering Intern

Carta — Waterloo, ON

May 2026 — Present

  • Automated the nightly refresh of Carta’s demo environment with a curated-snapshot pipeline — PostgreSQL dumps restored from S3/KMS via Kubernetes CronJobs and Helm — replacing 2.5 hours of manual re-seeding at 99.9% restore reliability.
  • Closed a cross-service authorization gap affecting 50+ services with an idempotent Spring Boot/Quartz job that replays access-control tuples, with least-privilege S3/KMS access provisioned through Terraform.
  • Built an agent-loop system that triages, fixes, and tests LLC account migrations, moving 100+ accounts with a 90% success rate and no human intervention.
  • Cut request latency 10% across eight database-heavy endpoints by decoupling them from a Django monolith with Kafka-based async event processing, and reduced CircleCI pipeline runtime by a third.

Software Engineering Intern

PengYang Asset Management — Fixed Income Desk, Beijing

May — Sept 2025

  • Built a Python engine simulating Black-Litterman (Idzorek) and All-Weather risk-parity portfolios, with daily runs completing in under a second.
  • Automated a SQL/Python ETL pipeline delivering daily risk data to traders at 99.9% uptime.
  • Deployed a predictive modeling module to production, backed by unit tests against historical market data.

University of Waterloo — BCS, Co-op

Sept 2024 — Present

Projects

Best Audio AI Hack · CxC 2026 · 350+ Teams

Deepfake-audio detector: a CNN trained on ASVspoof 2019 with real-time recording and a per-second confidence heatmap. Checkpoints published on Hugging Face.

99.8% test accuracy

Python · FastAPI · Next.js · PyTorch · Three.js

A from-scratch Llama-3 inference runtime with zero framework dependencies. INT4 quantization with a fused de-quant kernel fits a 120B-parameter model in 98 GB of VRAM on an RTX 6000.

Paged KV-caching · linear arena allocator

C++20 · CUDA · CMake

AlphaZero-style engine: an 18-layer ResNet trained purely on self-play, gaining over 1000 ELO across 2000 iterations. MCTS parallelized across 128 CPU cores for 4x search throughput.

5,000+ self-play games/hour

Python · PyTorch · Multiprocessing

Skills

Languages
C++ · CUDA · Python · Java · TypeScript · SQL
Systems & Performance
Multithreading · OpenMP · SIMD · INT4/8 Quantization · Memory Management · CMake · GDB · Valgrind
Machine Learning
PyTorch · Transformers · Hugging Face · NumPy · pandas · scikit-learn
Backend & Infrastructure
FastAPI · Django · Spring Boot · Kafka · PostgreSQL · Docker · React · Node.js · Git · Linux · CI/CD

Contact

I’m always open to talking about systems, ML infrastructure, and internship opportunities.

sumsjerrywu@gmail.com