Nam Le
AI Engineer · Sydney, NSW
I build something almost every day to find where AI coding agents break, then I read what they wrote.
Open to AI/ML engineering roles. Full Australian work rights (485 visa), no sponsorship needed.
Experience
DeliBI: a multi-tenant Vietnamese text-to-SQL analytics assistant (pre-launch). 223 commits.
- Built a plan → validate → execute → revise reasoning pipeline for the SQL analyzer, with self-correction on failed queries.
- Benchmarked 5 NL2SQL systems on 55 golden cases + 8 semantic traps. DeliBI v2 ranked first (63.3% SQL match vs 24.5% for v1); my report set the v2 direction.
- Shipped role-based PII masking with a bilingual VI/EN classifier.
- Made accent-insensitive search 5.6–8.8× faster with a Postgres generated column; cut embedding indexing from 41.3 s to 3.3 s.
- Owned VietQR payments, pricing tiers and quotas end to end.
- Led the root-cause analysis of a 457-duplicate-email incident; fixed it with dedup, guards and regression tests.
Selected projects
Each project says how it was built: by hand, mixed, or mostly by AI coding agents under my direction.
National Runner-Up and People's Choice, MentorME Futura Remix 2026.
- Candidates build with AI inside a browser sandbox (WebContainers, $0 VM cost) and are assessed on evidence, with a human mentor signing off high-stakes scores. Built ~130 of ~140 commits.
- Records and live-captions Vietnamese + English and summarises with a local LLM, no cloud. Cut English live-caption WER from 92% to 37% on code-switched speech; a 30-minute meeting processes end to end in 82 s.
- A compiler plugin + browser runtime that tells AI agents which element is covered, clipped or overflowing, with the exact
file:line:col. 34-case conformance corpus with 0 false positives / 0 false negatives. Open source (MIT); demo at an Atlassian-hosted event, 26 Nov 2026.
- 3D CNN + Bi-GRU + Conformer model at 89% accuracy; shrank the serving image 82% (2.8 GB → 450 MB). Selected for the UTS FEIT AI Showcase 2026.
- LLaMA-style 1M–49M parameter models (GQA, RoPE, SwiGLU) with a custom 2,048-token BPE; eval loss 4.55 → 2.38 in 1,000 steps on a consumer GPU.
- EfficientNet-B0 at 96.4% accuracy; INT8 quantization cut inference from 38 ms to 6 ms.
More: ReUnite (BLE mesh chat, Sync Hack USyd), Renova/ECOVAL (Innostar 2026 Grand Prize), Mai (local AI companion), go-serve-http (hand-written HTTP server). All repos on GitHub.
Education
WAM 78.81. High Distinctions in Cloud Computing (98), Databases (98), Image Processing (95), Deep Learning (93), AI Capstone (90).
Skills
- Languages
- Python, TypeScript, Rust, SQL, C++
- AI
- LLM agents and tool loops, RAG, evals and benchmarks, structured output, PyTorch, speech (ASR), computer vision
- Systems
- FastAPI, PostgreSQL, Tauri, Electron, Docker, GitHub Actions
Outside work
Hackathons (about 10 in 2026), photography, chess, nvim and tmux, ice skating, volleyball.