Backend and AI-infrastructure engineering — routing, retrieval, and gating expensive compute behind cheap fast paths.

Bachelor of Artificial Intelligence, University of Technology Sydney (expected 2028, GPA 3.7/4.0) · Backend AI Engineer — Internship (Remote) at FlyRank AI since July 2026 · Ho Chi Minh City

The thread through my projects: put a cheap, fast decision in front of an expensive, slow one. In Pulsemind (team), an XGBoost classifier's forward pass scores an ICU stream event in under 5 ms, and only anomalies go on to an LLM call. In Develarper (team of four), a local Qwen 2.5 3B model reads each task's difficulty and routes it before a cloud API is touched.

Portfolio · LinkedIn · X

A two-tier router built by team Develarper for the AMD Developer Hackathon (Act II). A local Qwen 2.5 3B model scores each task's difficulty; hard tasks go to a cloud model on Fireworks and easy ones stay local.

• 18 of 19 tasks routed to the correct tier, on a 19-task harness with self-defined labels.
• ~80% of tasks ran on the local model — a count of tasks, not tokens.
• 93.00% (186/200) answer quality over 100 factual and 100 summarization tasks, scored by similarity and Jaccard overlap against expected answers rather than human or LLM review. 6 of the 200 prompts contain unrendered template placeholders.
• 505.1 ms mean latency per task, across 200 sequential tasks.

_{Team of 4 — LLMOps · July 2026 · AMD Developer Hackathon, Act II · Python · Qwen 2.5 3B via Ollama · FastAPI · Fireworks}
Repository →

An asynchronous ICU telemetry processor. An XGBoost classifier, selected in a bake-off against LightGBM and CatBoost over 109 ventilator, comorbidity and treatment features, scores incoming stream events and gates higher-cost LLM rationalisation so it fires only on detected anomalies.

• < 5 ms classifier forward pass per stream event.
• < 50 ms end to end; the anomaly path adds the LLM rationalisation step, roughly 15 seconds.

_{Trained on the credentialed MIMIC-IV (PhysioNet) de-identified ICU dataset under its data use agreement; demonstrations run on a synthetic derivative. No clinical deployment.}

_{Team — full-stack AI engineer · May 2026 – ongoing · Python · XGBoost · PyTorch · scikit-learn · React}
UI prototype →

Built with team Hackaphobia at the GDGoC National Hackathon 2026 in Hanoi, where the team received a Top 30 Finalist Award. An onboarding survey and swipe interface: seven survey answers are encoded into a vector and scored against candidates by cosine similarity, with an optional free-text bio path through Vertex AI embeddings. Stateful WebSocket chat reaches every session a user has open.

• ~50 real users onboarded and swiped at demo day.
• ~12 ms average request latency on the structured matching path, excluding embedding generation.

_{Contributor, team of 4 — DevOps & Backend Engineering · April 2026 · Python 3.11 · FastAPI on Cloud Run · Cloud Firestore · Firebase Auth · Vertex AI · React 19 · Leaflet · Docker · GitHub Actions}
Repository → · Demo →

• Weatherise — a weather-intelligence pipeline: REST sources and NVIDIA Earth-2 surrogate models feed a multi-agent chain on Nemotron Ultra using MCP, with a Qdrant vector database for retrieval-augmented generation.
  _{Team — LLMOps & Backend AI Engineering · 9–11 June 2026 · Vietnam AI Open Hackathon (NVIDIA / OpenACC) · Python · Qdrant · Earth-2 · Nemotron Ultra · MCP}
• Vora — a quiz-verified learning roadmap. The backend resolver turns unstructured model output into validated, dependency-mapped directed acyclic graphs and checks prerequisite order. Backend repository.
  _{Team — backend engineer · April 2026 · GDGoC DevCamp, HCMUT · Express 5 · MongoDB · Zod · Fireworks · React 19 · React Flow}
• llmops — a self-hosted inference node on one machine: flan-t5-base in 8-bit with a LoRA adapter behind FastAPI, with PostgreSQL, Redis rate limiting, Traefik, a Cloudflare tunnel, and Prometheus and Grafana, all on Docker Compose.
  _{Solo · March 2026 · Python · FastAPI · PostgreSQL · Redis · Traefik · Prometheus · Grafana · Docker Compose}
• ArchitectureLab — a person and an agent inspect the same live system model; the agent calls tools the page registers through WebMCP and can only propose changes that a human applies. Repository.
  _{Team of 3 — scaffold, WebMCP adapter, deployment · September 2026 · OpenAI WebMCP Challenge · React 19 · Vite · TypeScript · Playwright}
• Local RAG API — a local-first RAG backend: ingest PDF or TXT, then chat over it. Chunks are embedded with all-MiniLM-L6-v2 into ChromaDB and answered through Ollama, and every chat response returns its sources and inference time.
  _{Solo · February 2026 · Python · FastAPI · ChromaDB · sentence-transformers · Ollama · MIT}
• agentrisk-daas — a risk-data service for the AI-agent package supply chain (MCP, npm, PyPI): a scraper on a six-hourly cron feeds a FastAPI, PostgreSQL and Redis API on Render, with a static Next.js front end. Site.
  _{Solo · March–August 2026 · Python · FastAPI · SQLAlchemy · Alembic · PostgreSQL · Redis · Next.js · Render}

• Languages — Python · TypeScript · SQL · Java
• Backend — FastAPI · REST APIs · WebSockets · async patterns · Cloud Firestore · PostgreSQL · Redis · SQLAlchemy / Alembic · Traefik
• AI infrastructure — XGBoost · PyTorch · scikit-learn · local LLM serving with Ollama · Vertex AI embeddings · Qdrant and RAG · multi-agent chains on Nemotron Ultra · NVIDIA Earth-2 surrogates · MCP
• Frontend — React 19 · Next.js 15 (App Router) · Tailwind v4
• Tooling — Git · GitHub Actions · Docker and Docker Compose · Prometheus and Grafana · Vercel

_{Regenerated daily by a GitHub Action in this repository. Top languages counts bytes of code in public, non-fork repositories.}

• Local RAG API starter kit — the paid companion to the open-source local-rag-api core described above.
• Next.js Mobile Starter Kit — built on the open-source nextjs-marketplace-free UI: Next.js 16 and React 19, a hand-rolled touch swipe deck, a 7-step onboarding wizard, per-user localStorage persistence and Radix-based primitives.

Portfolio · LinkedIn · X · YouTube · TikTok · Instagram