A cheap decision placed in front of an expensive one.
An XGBoost classifier scores ICU stream events 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. Same idea twice, and it is the part of this work I would defend in an interview.
Every figure below is stated with the conditions it was measured under. Where a project has no number, it is because nothing was instrumented — not because the number was disappointing.
|
🩺 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. Team — full-stack AI engineerMay 2026 — ongoingPython · XGBoost · PyTorch · scikit-learn · ReactClassifier forward pass <5 ms per stream event. End-to-end <50 ms; the anomaly path additionally incurs the ~15 s LLM step. Trained on the credentialed MIMIC-IV (PhysioNet) de-identified ICU dataset under its DUA; demos run on a synthetic derivative. No clinical deployment. |
🧠 Develarper LLM task routing A two-tier router built solo for the AMD Developer Hackathon (Act II). A local Qwen 2.5 3B classifier scores task difficulty and dispatches hard tasks to Fireworks while easy ones stay local. Solo — LLMOpsJuly 2026 · AMD Developer Hackathon, Act IIPython · Qwen 2.5 3B via Ollama · FastAPI · FireworksRouting: 18 of 19 tasks to the correct tier, on self-defined labels. Answer quality: 93.00% (186/200) over 100 factual + 100 summarization tasks, scored by similarity/Jaccard overlap against expected answers — not human or LLM review. Mean latency 505.1 ms, sequential. 6 of the 200 prompts contain unrendered template placeholders and are unanswerable as posed. Raw results are committed in the repo. |
|
⚡ Roomie Roommate and apartment matching Onboarding survey and swipe interface matching on structured fields, with an optional free-text bio path using Vertex AI embeddings and cosine similarity. Contributor, team of 4 — DevOps & Backend EngineeringApril 2026 · GDGoC National Hackathon 2026 (Hanoi), team HackaphobiaPython · FastAPI · React · PostgreSQL · Vertex AI~50 real users onboarded and swiped at demo day. ~12 ms average request latency — structured matching path only, excluding embedding generation. |
🌩️ Weatherise Weather-intelligence pipeline REST sources and NVIDIA Earth-2 surrogate models feeding a multi-agent chain on Nemotron Ultra, with a Qdrant vector database supporting RAG over the ingested data. Team — LLMOps & Backend AI Engineering9–11 June 2026 · Vietnam AI Open Hackathon (NVIDIA / OpenACC)Python · Qdrant · Earth-2 · Nemotron Ultra · MCPDemonstrated at concurrency 1 over three days. No latency or throughput figure is stated, because none was measured under load. |
|
🗺️ Vora Quiz-verified learning roadmap A deployed learning-roadmap application with quiz-gated progression. The backend resolver converts unstructured model output into validated, dependency-mapped directed acyclic graphs. Team — backend engineerApril 2026 · GDGoC DevCamp, HCMUTReact FlowCycle detection is O(V+E) — an algorithmic property of the traversal, not a measured latency. |
🧪 ArchitectureLab WebMCP architecture studio A shared live system model where a person and an agent inspect the same request flow. The agent calls structured tools the page registers and cannot change anything without approval. SoloAugust 2026 · OpenAI WebMCP ChallengeEvery number in the simulator is synthetic and directional — stated assumptions, not measurements. |
| Credential | Issuer | Type | Issued |
|---|---|---|---|
| Dean's List 2026 | University of Technology Sydney | Honour | 2026-07-09 |
| Certificate of Attendance | Vietnam AI Open Hackathon (NVIDIA / OpenACC) | Attendance | 2026-06-09 |
| Top 30 Finalist Award | GDGoC National Hackathon 2026 (Hanoi) | Placement | 2026-05-20 |
| Next.js App Router Fundamentals | Vercel | Completion | 2026-02-20 |
| Generative AI with Large Language Models | DeepLearning.AI & AWS — via Coursera | Completion | 2026-01-02 |
| AWS Cloud Practitioner Essentials | Amazon Web Services | Foundational completion | 2025-12-03 |
On the NVIDIA entry: the certificate is one of attendance. The claimable part is that the team was selected as one of 10 from roughly 100 registrants to compete. It was a one-round event and the team did not place.
- 📦 Local RAG API — a self-hosted retrieval-augmented generation API layer that runs on your own hardware.
- ⚡ Next.js Mobile Starter Kit — a Next.js starter configured for mobile-first static export.
- 📊 Global Tech Intelligence Node — a curated weekly feed of technology and infrastructure signals.