Quick to learn.
Hard to fool.
I know how I learn, so I adapt to new fields fast. I know how I fail, so I check every result before I trust it.
I learn at the speed
things change.
A few years ago, understanding a new field took me months. Then I started using AI not as an answer machine, but as a sparring partner: I ask it to explain, I test what it says against real sources, I find the gaps in my own understanding, and I go deeper right away. Topics that used to take months now take days.
That changed what I can work on. In the last year alone I've built models for satellite radar images, oil well sensors, intensive-care records, leaf photos, and macroeconomic data. Different fields, different data, different tools.
Learning fast is only half of it.
The other half is knowing where my thinking, and my models, can go wrong. So every project runs on the same four habits:
- Verify before trusting. A wrong number is worse than no number. I set honest baselines, lock how I'll evaluate before I see the results, and log every run so each number can be traced.
- Be honest about limits. Every project I show says what the evidence does not support. Sometimes the finding is that machine learning doesn't win, and that's worth knowing too.
- Keep learning. New field, new method, new tool. I treat unfamiliar problems as the normal case, not the exception.
- Share what I learn. I built a free site that teaches the maths behind AI in Indonesian, because explaining something is the best test of understanding it.
AI makes me fast. Computational thinking keeps me accurate.
Real problems, evaluated honestly.
Now
Why do ECG models fail on a new hospital's data?
An ECG model trained on German data performs worse on Chinese data. I'm designing a controlled study to separate the causes: label definitions, signal differences, or patient population. The protocol is written and pre-registered before the target data is opened.
Writing up the sepsis evaluation study
Turning the controlled leakage audit into a research paper. Results will be shared once it is published.
Lab
Skill Atlas
An interactive 3D star map of the AI agent skills I use and customise. Part tool, part toy, and a picture of how my AI setup fits together.
How I work with AI
I run research and development through coding agents with multi-agent workflows: structured method reviews, citation checks, and experiment logs. AI does the heavy lifting. I decide what to trust.
Experience
Research Assistant
Supporting several faculty research projects in data science and applied machine learning. Managing a grant-funded study on oil palm leaf disease detection with partners in Malaysia, and contributing to satellite-based mangrove mapping.
Data Analyst Intern
Environmental data analysis and ML models for water and air quality, field sampling, and public environmental education content.
Education
Skills
- Machine learning & deep learning scikit-learn, Random Forest, XGBoost, LightGBM, SVM, K-Means, PyTorch, CNNs, U-Net, DeepLabV3+, transfer learning, Grad-CAM
- Deployment & edge AI ONNX, INT8 quantisation, in-browser inference (onnxruntime-web), Streamlit apps, live monitoring dashboards
- Evaluation & research methods leakage audits, spatial cross-validation, pre-registration, reproducible pipelines, experiment logging
- Statistics, time series & econometrics hypothesis testing, bootstrap, SARIMAX, LSTM, GARCH-MIDAS, CUSUM, Diebold-Mariano, McNemar, Friedman and Cochran's Q tests, statsmodels
- Geospatial & remote sensing Google Earth Engine and GEE apps, Sentinel-1 SAR, Sentinel-2, Landsat, spectral indices (NDVI, NDWI, MNDWI), rasterio/GeoTIFF
- AI engineering Claude Code, Codex, multi-agent workflows, RAG, prompt and context design, AI-generated video (this site's background)
- Data, web & visualisation pandas, NumPy, Matplotlib, Seaborn, Plotly, Tableau, Excel, JavaScript, Vite, Vercel, Git & GitHub
- Writing & teamwork journal manuscripts and LaTeX, technical documentation, team lead of a six-person deep learning capstone, environmental field sampling
Let's work on something
hard.
Internships, research collaborations, or a problem you're stuck on. I usually reply within a couple of days.