Applied AI, shown through the decisions it changed.
Selected systems from research and production—each documented through the problem, intervention, evidence, and remaining limitation.
Three systems. Three kinds of reliability.
Safety, multimodal perception, and cultural evaluation—shown at full resolution.
Do Multilingual VLMs Abstain Under Cross-Modal Conflict in Low-Resource Languages?
A model that trusts visual evidence in English follows a false caption in Telugu. We measured the shift, located its signal, and steered the model back.
- Problem
- Vision-language models must choose between visual evidence and text when the two conflict. That choice matters for safety. Hundreds of millions of people use Hindi, Telugu, and other languages with limited training and red-team...
- Intervention
- We built and released four conflict datasets in English, Hindi, and Telugu. They cover natural scenes, physics renders, remote sensing, and rendered 3D objects. NLLB-200-600M and Claude produced translations that people then checked. Each...
- Evidence
- This is the first multilingual audit of cross-modal conflict that controls for visual perception. It shows where grounding fails beyond English. The override gap lets deployers and regulators identify the...
- Limit / status
- I will first test how far an honesty direction learned in one language transfers across many others and identify where it breaks. I will also develop a standard,...
Hidden-State Detection of In-Context Goal Hijacking with a Conformal False-Positive Guarantee
Output monitoring misses failed hijack attempts. The model's residual stream detects almost all of them.
- Problem
- Most deployed LLM security tools monitor inputs and outputs. A goal-hijack attack hides an instruction in the prompt to override the user's task. It changes the model's internal computation even when the final answer...
- Intervention
- I built a read-only pipeline for Qwen2.5-0.5B-Instruct and ran one forward pass on each of 640 prompts. The four groups were clean prompts, attacks, harmless prefixes using attack vocabulary in safe ways, and an...
- Evidence
- On one model family and attack style, a clean-versus-attack benchmark gave perfect AUC to a detector that flagged every harmless prefix. Calibration on one benign format also promised a 5%...
- Limit / status
- With hard negatives, the 0.5B layer sweep rises from chance at the embeddings, peaks in the early-middle layers, then fades near the output head. That pattern disappears at...
Flood segmentation from multispectral satellite imagery (ProCANet)
State-of-the-art flood maps for a basin of 25 million people, produced within hours.
- Problem
- Disaster teams need flood maps before they can plan routes and supplies. Manual GIS analysis takes days or weeks, leaving little time in the first 72 hours of an emergency. The 270 km Citarum...
- Intervention
- I designed ProCANet, a dual-encoder U-Net with separate RGB and NIR branches. Cross-attention at every skip connection combines both inputs at four scales: 1/2, 1/4, 1/8, and 1/16. Self-attention first captures long-range patterns within...
- Evidence
- The system can support disaster response for more than 25 million people in the Citarum basin. It cuts flood mapping from days of manual GIS work to hours after a...
- Limit / status
- Documented results and constraints are available in the full case.
Every case, one clear entry point
I update these case studies as the work develops. Each one explains the problem, method, evidence, and limits so others can assess or reproduce it.
Flood segmentation from multispectral satellite imagery (ProCANet)
State-of-the-art flood maps for a basin of 25 million people, produced within hours.
Mining footprint detection with multi-modal satellite data
The first published use of a geospatial foundation model to map mining footprints at continental scale.
Aquaculture pond detection and change analysis
Detecting illegal shrimp farms from orbit within hours of a satellite pass.
Flood policy evaluation, retention pond effectiveness in South Bandung
We used satellite data and causal analysis to test whether retention ponds reduce flooding.
Do Multilingual VLMs Abstain Under Cross-Modal Conflict in Low-Resource Languages?
A model that trusts visual evidence in English follows a false caption in Telugu. We measured the shift, located its signal, and steered the model back.
Hidden-State Detection of In-Context Goal Hijacking with a Conformal False-Positive Guarantee
Output monitoring misses failed hijack attempts. The model's residual stream detects almost all of them.
SEA-VL: multicultural vision-language benchmark for Southeast Asia
We built the ACL 2025 benchmark that measures GPT-4V's failures on Southeast Asian culture.
GG-EZ: regional adaptation framework for vision-language models in SEA
Adapted image generation for Southeast Asia while retaining more than 98% of global benchmark quality.
CommonLID: language identification on noisy web data
Language identification scores above 95% on clean tests fall sharply on real Southeast Asian web data.
Share of Voice forecasting system, Fortune 500 APAC (Artefact)
Replaced media guesswork with calibrated forecasts across six APAC markets for a Fortune 500 company.
Biometric authentication and alternative credit scoring at scale (GDP Labs)
More than one million daily inferences at 99.99% uptime, plus credit scores for borrowers with little formal history.
Real-time fraud detection pipeline (GDP Labs)
One real-time pipeline detected account takeover, synthetic identities, and merchant collusion.
Municipal waste logistics forecasting, Jakarta Smart City
Forecasted Jakarta's waste so collection routes could follow demand throughout the week.
Demand and audience forecasting for financial and media clients (Artefact)
Delivered two forecasting systems for Fortune 500 clients in two-week consulting sprints.
HakkTaxi: ride-share demand prediction (Microsoft Azure APAC Hackathon)
Built a Jakarta ride-demand heatmap in 48 hours and won the Microsoft Azure APAC regional championship.
TeleHealthMonitor: Edge AI for remote patient monitoring (CamvsCovid)
Monitored COVID-19 vital signs over 2G without sending video to the cloud; placed in Cambridge's global top three.
Community IVR: Voice AI for offline communities (Cal Hacks 8.0)
Delivered health information by phone to Indonesians without smartphones or reliable internet.
Plastic bag ban, causal policy analysis using NLP and citizen data
Used more than 100,000 citizen complaints to measure the effect of Jakarta's 2020 plastic bag ban.
Qwen VL Fine-tuning for AI City Challenge 2026 Track 2
Fine-tuned a three-billion-parameter vision-language model on one 14 GB GPU.
ENCP: Episode-Normalized Conformal Prediction for Vision-and-Language Navigation
A parameter-free rescaling turns an overconfident navigation policy into a calibrated signal for when to ask for help.
Contributing to llm-d/inference-scheduler, Kubernetes LLM Inference Scheduling
Contributing to the Kubernetes scheduler that routes language-model requests across GPU backends.
Batch-to-Streaming Energy Forecasting with Spark ML and Kafka
Trained a building-energy model in Spark, then used it to score a live Kafka weather stream.