Work in Progress

This is a living document. Use cases are documented progressively as part of an ongoing effort to make applied work accessible and reproducible. Check back for updates.

Semantic SegmentationAttention MechanismsMulti-Modal FusionRemote SensingPyTorch Lightning

Flood segmentation from multispectral satellite imagery (ProCANet)

IEEE Geoscience and Remote Sensing Letters (GRSL), 2025 -- Q1 (IF 4.4) · First Author

Satellite flood maps for 25M people — hours instead of days, state-of-the-art IoU.

PyTorchPyTorch LightningGoogle Earth EngineSentinel-2PlanetScopeWeights & Biases (W&B)GeoTIFFmultispectral rasterNDWI
Read writeup →
Multi-Modal FusionSemantic SegmentationSARDEMFoundation ModelsLand Use

Mining footprint detection with multi-modal satellite data

Remote Sensing of Environment, vol. 318, 2025 -- Q1 (IF 11.4) · Co-Author

The first geospatial foundation model applied to continental-scale mining footprint detection.

TensorFlowPyTorchGoogle Earth EngineSentinel-1Sentinel-2DEMPrithvi (NASA/IBM)
Read writeup →
Object DetectionCoastal MonitoringChange DetectionTime-Series SatellitePatent

Aquaculture pond detection and change analysis

Monash University -- Filed Patent IDS000010594, Issued June 2025 · Lead Developer

Catching illegal shrimp farms from orbit before regulators arrive — hours, not field-survey years.

PythonGoogle Earth EngineNDWIMNDWIAWEISentinel-2 (multi-temporal)
Read writeup →
Vision-Language ModelsBenchmarkDatasetCrowdsourcingCultural AIACL 2025

SEA-VL: multicultural vision-language benchmark for Southeast Asia

ACL 2025 (Main Conference, Long Paper) · Core Contributor -- Data Pipeline Lead

GPT-4V fails on Southeast Asian culture. We built the benchmark that proves it — ACL 2025.

GPT-4VGemini 1.5Claude 3LLaVAInternVLpHashCLIP-ViTSigLIPNomic Embed Vision100+ annotators, 11 countries10,000+ image-question pairs
Read writeup →
ForecastingConformal PredictionXGBoostGCP Vertex AIdbtMLOpsUncertainty Quantification

Share of Voice forecasting system -- Fortune 500 APAC (Artefact)

Artefact -- Senior Data Scientist (Feb 2025 -- Nov 2025) · Lead ML Engineer (Founding Technical Member, Jakarta Office)

Turned media gut-feel into calibrated prediction intervals across 6 APAC markets for a Fortune 500.

XGBoostConformal Prediction (split)GCP Vertex AIBigQueryGCSCloud WorkstationdbtPythonSQLSimilarWebTraackrStreamlitLangChainLangSmithRAGSplit conformal predictioncalibration drift monitoring
Read writeup →
BiometricsFace VerificationAnti-SpoofingLiveness DetectionCredit ScoringFintechMLOpsRegulatory Compliance

Biometric authentication and alternative credit scoring at scale (GDP Labs)

GDP Labs (GLAIR.ai) -- Senior Data Scientist / ML Engineer (2021 -- 2023) · Lead ML Engineer -- Promoted to Senior within 12 months

1M+ daily inferences at 99.99% uptime. Credit for the borrowers banks wouldn't score.

PyTorchscikit-learnXGBoostIntel OpenVINOMobileNetMLflowDockerKubernetesAWS (S3EC2ECRSageMakerLambdaAPI Gateway)JMeter (sync + async)MySQLPrestoDBSHAPScorecard (logistic regression)
Read writeup →
Fraud DetectionAnomaly DetectionGraph NetworksReal-Time InferenceKafkaRedisFintech

Real-time fraud detection pipeline (GDP Labs)

GDP Labs (GLAIR.ai) -- Lead ML Engineer (2021 -- 2023) · Lead ML Engineer

Three fraud attack types, one pipeline — rule filters, XGBoost, and graph anomaly detection.

XGBoostgradient boostinggraph anomaly detectionApache KafkaRedisPythonscikit-learnDockerKubernetesAWS
Read writeup →
Time Series ForecastingSmart CityProphetLSTMCausal InferencePolicy Analysis

Municipal waste logistics forecasting -- Jakarta Smart City

Jakarta Smart City -- Data Scientist (Jan 2021 -- June 2021) · Lead Data Scientist

Forecasted Jakarta's trash so trucks go where waste actually is, not where the schedule says.

PythonRFacebook ProphetARIMASARIMALSTMstatsmodelsTableauMatplotlibscikit-learnpandasnumpyDifference-in-differenceshypothesis testing
Read writeup →
Gradient BoostingGeospatial ClusteringDemand PredictionAzure MLReal-TimeHackathon

HakkTaxi: ride-share demand prediction (Microsoft Azure APAC Hackathon)

Microsoft Azure Virtual Hackathon APAC -- Regional Champion (2020) · Lead ML Engineer

A Jakarta ride-demand heatmap built in 48 hours. Microsoft Azure APAC Regional Champion.

Microsoft Azure (Azure MLAzure Maps)XGBoostgradient boostingH3 gridPOI density featuresHistorical trip logsweather APIPOI data
Read writeup →
Edge AIHealthcareComputer VisionGDPR ComplianceLow BandwidthHackathon

TeleHealthMonitor: Edge AI for remote patient monitoring (CamvsCovid)

CamvsCovid, Cambridge Judge Business School -- Top 3 Globally (2020) · Lead ML Engineer

COVID-19 vitals monitoring on a 2G connection — no cloud, no smartphone. Top 3, Cambridge.

On-device (ARM CPU)optical flowpose estimationModel quantization (INT8/FP16)OpenCV2G/3G compatible (minimal data transmission)PythonOpenCVTensorFlow Lite
Read writeup →
NLPSpeech RecognitionIVRAccessibilityLow-ResourceHackathon

Community IVR: Voice AI for offline communities (Cal Hacks 8.0)

Cal Hacks 8.0, UC Berkeley -- Best Community Track (2020) · Lead ML Engineer

Health info for 40% of Indonesians without smartphones — via a plain old phone call.

ASRTTSDTMF IVRKnowledge graphintent classificationPSTN (standard telephone network)2G compatibleIndonesian + Javanese/Sundanese dialect variants
Read writeup →
NLPCausal InferenceDifference-in-DifferencesPolicy AnalysisIEEESmart City

Plastic bag ban -- causal policy analysis using NLP and citizen data

ICISS 2021 -- IEEE · First Author

Jakarta banned plastic bags in 2020. Did it work? Proved it causally with 100K+ complaints.

Text classificationcomplaint taxonomyDifference-in-differenceshypothesis testingJAKIQlue citizen complaint platformsPython statsmodelsR100,000+
Read writeup →
Vision-Language ModelsLoRA Fine-tuningQwen2.5-VLLLaMA-FactoryDeepSpeedVLM TrainingECCV 2026

Qwen VL Fine-tuning for AI City Challenge 2026 Track 2

AI City Challenge 2026 Track 2 -- ECCV Workshop (2026) · Lead ML Engineer

Fine-tuning a 3B VLM on a 14GB GPU — every memory constraint hit, diagnosed, and solved.

Qwen2.5-VL-3B-InstructLoRALLaMA-FactoryDeepSpeed ZeRO-2PyTorch 2.5.1+cu121pefttransformersV100 14GBCUDA 12.5
Read writeup →
Conformal PredictionVision-Language NavigationUncertainty QuantificationVLN-DUETVLN-HAMTScore FunctionsICRA 2027

Conformal Prediction for Vision-and-Language Navigation (DUET + HAMT)

Thesis Research -- VLN-DUET Uncertainty Propagation (2026) · Lead Researcher

21 uncertainty methods for robot navigation tested. Most fail. One works — and we know why.

PyTorchDeepSpeed ZeRO-2VLN-DUETVLN-HAMTTHRAPSRAPSConformal Prediction21 formulations (EMA, CumulMax, ACI, PID, entropy-family, etc.)R2RR4RREVERIESOONCoverage guaranteesSet sizeIntervention rateA40V100
Read writeup →
Open Source ContributionKubernetesGoInference SchedulingGateway APIEnvoyP/D DisaggregationKV Cache

Contributing to llm-d/inference-scheduler -- Kubernetes LLM Inference Scheduling

Open Source -- llm-d/inference-scheduler (Planned 2026) · Contributor

Contributing to the Kubernetes layer routing LLM requests across GPU backends at production scale.

Go 1.24+Gateway APIEnvoy ext-procKustomizeHelmvLLMP/D DisaggregationKV CachePrometheusGrafanaDistributed tracingUnit testsIntegration testsE2E (Kind clusters)TinyLlama-1.1BQwen3-VL-2B
Read writeup →