2023 — present
Autonomous Water Treatment
A reinforcement learning and agentic AI framework that runs ultrafiltration plants end to end, from raw telemetry to verified control setpoints.
My doctoral research, conducted under the U.S. Army Corps of Engineers, asks a blunt question: can a water treatment plant run itself well enough to beat a skilled human operator on net water production, without ever violating a safety constraint?
The framework has three layers.
1. Perception — forecasting the plant
Ultrafiltration telemetry is non-stationary by nature: membranes foul, feed water quality drifts seasonally, and cleaning events reset the system to a different operating regime. A single global model averages across all of that and predicts poorly at exactly the moments that matter.
The pipeline instead uses late-fusion forecasters over separate sensor streams, with regime decomposition applied first to isolate distinct historical operating states. Anomaly monitoring runs alongside on the operational KPIs, so a forecast that is confidently wrong gets flagged rather than acted on.
2. Policy — reinforcement learning for control
Backwash and cleaning scheduling is the lever with the largest effect on net water production, and it is conventionally handled by fixed rules. I train deep RL policies over a plant simulator calibrated to real telemetry, with the reward tied directly to net production and hard constraints encoded as terminal penalties rather than soft costs.
Offline policy evaluation gates anything before it touches hardware. A policy that cannot be shown to beat the incumbent rule-based schedule on historical data does not get deployed.
3. Action — an agentic controller with guardrails
The top layer turns forecasts into setpoints. It is an agentic controller, but the interesting engineering is in what sits between the model and the actuator: a constraint verification layer that checks every proposed setpoint against the plant’s operating envelope and rejects, rather than clips, anything outside it.
Current direction
I am pretraining water treatment foundation models on multi-plant telemetry and post-training them per site, so a new plant inherits a strong prior instead of starting from a cold model with three months of data.