The hard kind of agentic. Sensors in. Regulator-readable reports out.
Our flagship engagement is a pod embedded with WaterDoctor, an NUS spinoff building AI-integrated biofilm reactors for water and aquaculture — and a company wGrow co-invested in. The crew turns sensor and lab telemetry into adaptive treatment loops, with run summaries an environmental regulator can sign off. Four agents, two senior wGrow engineers, weekly release.
Streams from in-tank probes (DO, pH, ammonia, nitrate, turbidity) and weekly lab assays. Normalises units, marks suspicious readings, archives raw.
Decides recirculation rates, aeration, biofilm cycles based on a model trained on the client's process. Bounded by safety envelopes the engineers wrote.
Generates weekly run reports the client uses with regulators. Every claim is linked back to raw data; nothing is paraphrased.
Code review every controller change. Own the safety envelopes. Sign every regulator-bound report.
- The agents have narrow jobs. No "do everything" agent. Each one is testable.
- The control loop has bounds. The agent picks moves; the engineers wrote the floor and ceiling.
- Every report is grounded. No agent text reaches a regulator without trace-back to raw data.
- Eval is continuous. Deviation from the predicted curve is flagged before the next dose.
Reported outcomes from the client (industrial pilots): 50%+ pollutant discharge reduction, 90% lower water exchange, 30% lower energy per unit treated.
Read the WaterDoctor engagement →