Offline-first early warning for hospital wards. In simulation, it caught 56% of deteriorations in time, compared with 36% for NEWS2 at a similar alert rate.
One screen per nurse or physician showing which patient to see first, ranked by NEWS2 with each patient’s own baseline breaking ties. Built to work with no network: chest patch, ward gateway, edge server, cloud sync when it comes back. The simulation used a distilled MindSpore model.
Simulator results, not clinical results.
- MindSpore
- Edge
- NEWS2
- Offline sync
In progress
Flood risk, drawn on the map you’re looking at.
Predicts flood risk for the area on screen from live NOAA precipitation forecasts and local terrain. An ANUGA physics simulation does the real work; a PyTorch U-Net is the fallback when it can’t. Results land as colour-coded heatmaps.
- FastAPI
- React
- TypeScript
- MapLibre GL
- PyTorch
- ANUGA
Innovation of Service
The energy anatomy of transformer pretraining: where does the energy go?
Our framework compares each component’s share of measured energy with its share of compute. The dashboard puts those measurements side by side across 18 configurations, so you can see which parts merit a closer look.
I led the research and co-authored it with Bill Ian Alcantara. We profiled three Pythia model sizes on an NVIDIA T4.
- GPU profiling
- Roofline model
- Dashboard
APCORE I-RISE winner
A search for “robin” might return “bird” or “nest.” Can the system tell how each relates?
We proposed the Semantic Depth Index to evaluate what vector retrieval preserves about meaning. It scores how concepts cluster, whether categories hold together, whether analogies carry over, and how closely the results match human judgments.
I co-developed and presented the framework. It won Best Presentation and Best Undergraduate Research at APCORE I-RISE in 2025.
- Vector retrieval
- Embeddings
- Semantic Depth Index
APCORE I-RISE winner
Smart urban mobility for walking Makati.
Hazard-aware walking routes, geotagged AR hazard reports, SOS video alerts and a city-side ticket dashboard. Built in 24 hours with the team that won the 2026 Poland–Philippines IT-Business Summit hackathon.
Grand Prize
A Chrome extension that scores the page you’re on.
Rates phishing risk across 11 URL, content and scam-language signals, and explains itself on the page. Language classification with a heuristic fallback, plus local and shared caching so repeat checks don’t re-pay the inference latency.
- JavaScript
- FastAPI
- QStash
- Supabase
Reusable agent skills, written once and run everywhere.
In BioSentry, my small-plan skill turns a change into a brief with the files to touch and a check to run. A lighter agent handles the repetitive edit and reports the test result for review.
- Agents
- Claude Code
- Workflows
Small changes that bend the tools to how I work.
Activity shows which sessions and tasks are running. Decisions collects open questions and sends my answer back to the coordinating agent. Folds saves long tool output to files and leaves a short note for the next session.
Nothing under that tag yet.