Systems Engineering // Core Engine
A cheap fast path,
deciding whether the expensive one runs.
Two of the projects below are the same idea twice. An XGBoost classifier’s forward pass scores an ICU stream event in under 5 ms and only then pays for an LLM call. A local Qwen 2.5 3B model reads a task’s difficulty and routes it before a cloud API is touched. Put a cheap decision in front of an expensive one and the expensive one stops setting the pace.
Current Experience
Model integration and API design. Work in progress.
Credentials
| Credential | Issuer | Type | Issued | Record |
|---|---|---|---|---|
Dean's List 2026 | University of Technology Sydney | Honour | 2026-07-09 | [VIEW_CRED__→] |
Certificate of Attendance | Vietnam AI Open Hackathon (NVIDIA / OpenACC) | Attendance | 2026-06-09 | [VIEW_CRED__→] |
Top 30 Finalist Award | GDGoC National Hackathon 2026 (Hanoi) | Placement | 2026-05-20 | [VIEW_CRED__→] |
Next.js App Router Fundamentals | Vercel | Completion | 2026-02-20 | [VIEW_CRED__→] |
Generative AI with Large Language Models | DeepLearning.AI & AWS — via Coursera | Completion | 2026-01-02 | [VIEW_CRED__→] |
AWS Cloud Practitioner Essentials | Amazon Web Services (AWS) | Completion | 2025-12-03 | [VIEW_CRED__→] |
Core Pillars Matrix
Tier 1 — Featured work
- TIER 1
Pulsemind
Critical-care telemetry — research prototype
An asynchronous ICU telemetry processor. An XGBoost classifier scores incoming stream events and gates higher-cost LLM rationalisation calls so they fire only on detected anomalies.
- Role
- Team — full-stack AI engineer
- Period
- May 2026 — ongoing
- Stack
- Python · XGBoost · PyTorch · scikit-learn · React
- Classifier forward pass
- <5 msper stream event
- TIER 1
Weatherise
Weather-intelligence pipeline
A weather-intelligence pipeline built at the Vietnam AI Open Hackathon (NVIDIA / OpenACC). REST data sources and NVIDIA Earth-2 surrogate models feed a multi-agent chain running on Nemotron Ultra, with a Qdrant vector database supporting retrieval-augmented generation over the ingested data.
- Role
- Team — LLMOps & Backend AI Engineering
- Period
- 9–11 June 2026
- Built at
- Vietnam AI Open Hackathon (NVIDIA / OpenACC)
- Stack
- Python · Qdrant · NVIDIA Earth-2 surrogates · Nemotron Ultra · MCP · REST sources
- TIER 1
Vora
Quiz-verified learning roadmap
A deployed learning-roadmap application with quiz-gated progression, built at the GDGoC DevCamp at HCMUT — a separate programme from the national hackathon. The backend resolver converts unstructured model output into validated, dependency-mapped directed acyclic graphs.
- Role
- Team — backend engineer
- Period
- April 2026
- Built at
- GDGoC DevCamp, HCMUT
- Stack
- Express 5 · MongoDB · Zod · React 19 · React Flow · Vite
Tier 2 — Further work
- TIER 2
Roomie
Roommate and apartment matching
A student roommate and apartment matcher built at the GDGoC National Hackathon 2026 in Hanoi with team Hackaphobia. An onboarding survey and swipe interface match on structured fields — location, budget — with an optional free-text bio path using Vertex AI embeddings and cosine similarity.
- Role
- Contributor, team of 4 — DevOps & Backend Engineering
- Period
- April 2026
- Built at
- GDGoC National Hackathon 2026 (Hanoi) — team Hackaphobia
- Stack
- Python · FastAPI · React · PostgreSQL · SQL · Vertex AI embeddings
- Users onboarded
- ~50real users who onboarded and swiped at demo day
- Average request latency
- ~12 msstructured matching path only — excludes embedding generation
- TIER 2
Develarper
LLM task routing
A two-tier LLM router built solo for the AMD Developer Hackathon (Act II). A local Qwen 2.5 3B classifier scores each task's difficulty and dispatches the hard ones to a cloud API while the easy ones stay on the local model. Tested against a 19-task harness.
- Role
- Solo — LLMOps
- Period
- July 2026
- Built at
- AMD Developer Hackathon, Act II
- Stack
- Python · Qwen 2.5 3B via Ollama · FastAPI · Fireworks
- Correct tier routing
- 18 of 19tasks, scored against self-defined labels
- Ran on the local model
- ~80%a TASK count, not tokens — the cloud-routed fifth were the hard tasks, carrying longer prompts and longer completions