Skip to content

Hi, I'm
Leo Nguyen

AI/ML Engineer & Researcher

I build agentic systems for both enterprise internal and customer-facing use cases. I also do research on agents verifiability, security, and governance.

View my work
Leo Nguyen, AI/ML Engineer & Researcher

Selected work

All 7 projects
L33t KV Store project screenshot
Systems2025 - 2026

L33t KV Store

A from-scratch study answering how fast a key-value store can actually go on commodity hardware once protocol and framework overhead are stripped away, browseable end-to-end so the answer is reproducible, not just claimed.

  • C
  • WebAssembly
  • Next.js
  • Engineered a custom binary-protocol KV store in 400 lines of C that sustains 36,000 ops/sec at 80 microsec latency on a 3-node LAN, beating Redis 6.0 throughput by 2%.
  • Drove server throughput locally from 25,000 to 185,000 ops/sec across 4 implementations spanning Python asyncio, uvloop, C with edge-triggered epoll, and C with io_uring.
  • Deployed the project on Vercel as a Next.js app that runs the C protocol parser in-browser via WebAssembly.
Live appSource
ResearchIn progressDecember 2025 – Present

Zero-Knowledge Federated Learning for Medical Imaging

Lets hospitals collaboratively train medical-imaging models without trusting the aggregating server: every server-side update arrives with a proof any participating site can verify in milliseconds.

  • PyTorch
  • Flower
  • Rust
  • EZKL
  • Nova Folding
  • GCP
  • 30KB aggregation proofs with 0.3ms verification — catches weighted-update tampering
  • Head-only proving on EfficientNet-B0: 5h → 59s proof generation (97% reduction); 200KB proofs, 4.6s verify
  • Reproducible Flower + Docker + Kubernetes + GCP stack; cut 112s setup overhead via artifact caching
NeuroScan project screenshot
MLOct 2024 - Nov 2024

NeuroScan

A clinical decision-support tool for radiologists doing brain-MRI triage: shows the predicted diagnosis alongside saliency maps and an LLM rationale, so a 99%-accurate model can't sneak past audit on hidden bias.

  • TensorFlow
  • Gemini
  • Next.js
  • Engineered a production deep-learning brain tumor classifier in TensorFlow that hit >99% test accuracy across many CNN architectures (Xception, ResNet50V2) on 7,023 MRI scans spanning 4 diagnostic classes.
  • Shipped a live app with Next.js and Vercel against a Dockerized FastAPI inference service on HuggingFace Spaces, delivering <3 seconds end-to-end predictions to users across 3 switchable models per request.
  • Built an explainability layer pairing TensorFlow GradientTape saliency maps with Gemini 2.5 multimodal LLM analysis to catch bias in 99%-accurate models that would have failed clinical audit despite high accuracy.
Live appSource

Right now I am architecting agentic systems at Canaan Group: RAG-backed agents that propose follow-ups, surface revenue, and flag anomalies inside a live CRM.

Read the background