AI for Physical Systems.
Understanding. Predicting. Optimising the real world.
We build AI systems that understand, predict and optimise physical systems, from industrial equipment and energy infrastructure to environmental sensing and autonomous machines.
- Research Directions
- 5
- Interactive Demos
- 4
Energy grids to factory floors.
We build where data meets the physical world. These are the problems we refused to solve the easy way.
Grid Energy Intelligence
Forecasting demand and peak load at state scale.
Multi-horizon energy demand and peak load forecasting with calibrated uncertainty across all 32 Indian states. The model captures seasonal patterns, industrial activity, and weather-driven consumption dynamics.
Pollution Intelligence
Predicting air quality at city scale.
Spatiotemporal models that learn from distributed sensor networks to forecast pollutant concentrations across urban and industrial regions, with calibrated uncertainty.
Battery Health Intelligence
Predicting degradation before capacity fades.
Electrochemical model hybrids that forecast battery degradation trajectory and remaining useful life from cycling data, temperature profiles and impedance spectra.
Predictive Maintenance
Hearing failure before it happens.
A multimodal sensor-fusion stack that forecasts machine failure windows from vibration, acoustic and thermal streams. Designed for industrial deployment and built for operational environments.
Four models. One approach.
Predictive maintenance, pollution forecasting, battery intelligence, energy forecasting. Different sensors, different time scales. Same core ideas running through all of them.
Sensor & Time-Series First
Every model we build starts from sensor streams: vibration, thermal, chemical, electrical, atmospheric. We do not assume clean, curated features. The architecture adapts to the data.
Hybrid Physics + ML
Pure black-box models extrapolate badly. Pure physics models are too expensive to run in production. So we embed domain constraints like electrochemical equations, atmospheric dynamics and fatigue mechanics into learned architectures that stay grounded in reality without becoming computationally impractical.
Calibrated Uncertainty
A point prediction that is wrong is worse than no prediction at all. Every model we build outputs calibrated uncertainty intervals that widen naturally where data is sparse or conditions shift. You know what the model does not know.
Built for Deployment
Notebook models do not ship. Every architecture is designed from the start for inference constraints: latency budgets, edge hardware, sensor dropout, distribution shift. Research that never leaves the notebook is incomplete.
Have a problem worth a thesis?
If your problem is hard enough to need genuine research and important enough to justify deployment, we would like to hear about it.