We build production AI systems for Bangladesh's hardest problems—geospatial intelligence, industrial automation, public health, and government services. Four platforms validated on $300 budgets, now scaling to national infrastructure.
Four production platforms built and deployed with minimal resources, achieving state-of-the-art performance through physics-informed learning, foundation model fine-tuning, and federated architectures.
SegFormer-B2 hierarchical vision transformer trained on physics-generated pseudo-labels. Fusion of Sentinel-1 SAR backscatter, Sentinel-2 NDWI, and SRTM slope constraints. Processing 2,847 km² scenes in 8 minutes on CPU—336× faster than manual GIS methods.
View Flood Demo →FATE 1.11 framework with FedAvg algorithm and Paillier homomorphic encryption. Raspberry Pi 4 edge devices run 5M-parameter LSTM models locally. Pilot across 5 garment factories achieved 85% failure prediction precision.
Explore Digital Twin →Combines Facebook Prophet with Tigramite's PCMCI algorithm for lag-specific causal discovery across 36 covariates. Outperforms univariate ARIMA by 32% and Prophet-only baseline by 18% on 7-day forecasts.
See Dengue Forecasting →Stacked ensemble of CatBoost, LightGBM, and XGBoost with Bayesian hyperparameter optimization. Engineered proprietary "Trade Imbalance Ratio" feature accounting for 38% of predictive power.
Analyze Freight Model →Real-time economic nowcasting using VIIRS satellite nightlights with advanced denoising and cloud removal algorithms.
View Nightlights →Zero-label crop stress discovery using self-supervised learning on multispectral imagery for early intervention.
Explore Crops →Time-series decomposition with XGBoost for accurate monthly LPG demand forecasting across distribution networks.
View LPG Model →Allen AI invested $2M training SatlasPretrain on 302M satellite images. We fine-tune their open-source Swin-v2-Base encoder on 10,000 Bangladesh-specific tiles for $150 in compute—achieving 15-20% accuracy gains on local tasks.
Our systems run on $50 Raspberry Pi devices via ONNX Runtime with INT8 quantization, achieving <200ms inference latency. Federated learning ensures 99% of computation happens on-premise.
Standard neural networks memorize patterns; physics-informed neural networks embed conservation laws directly into loss functions, forcing generalization to unseen conditions.
Bangladesh serves as our beachhead market with 170M population, $460B GDP, and critical climate vulnerability. Proven models scale to Nepal, Sri Lanka, and broader South Asian markets.
Scale HAWKEYE into multi-task foundation model serving government agencies. Target: $500K-2M annual contracts.
Onboard 100 garment factories to federated predictive maintenance. Target: $1M ARR at $10K/factory.
Deploy LayoutLMv3 document intelligence across ministries and hospitals. Digitize 10M records.
We have the technical proof, market understanding, and operational discipline. Seeking partners who recognize the strategic value of building frontier AI for emerging markets.
