Agent-driven dosing decisions for water treatment, with human supervision — out-of-range actions require human confirmation.
About this page
This page only depicts a conceptual architecture and capability summary. No customer production data, sensor readings, or live system UI is shown.
In short
FuGuang Smart Phosphorus Control is an agent-driven dosing design for water treatment. Sensor arrays feed water quality, flow and dosing figures into the agent context at second-level latency; the agent combines threshold models with LLM reasoning to propose a dose and a risk level; only after an operator confirms it does the actuator adjust, and the outcome flows back as an adaptive closed loop. The key difference from fixed PID control is that any decision outside the critical thresholds requires human confirmation and is fully audit-logged — the agent advises, it does not decide. This page shows architecture and capability only, with no customer production data.
What the agent takes off people's plates
Sensor arrays feed water quality, flow, and dosing metrics into the agent context at second-level latency.
Threshold models combined with LLM reasoning produce dosing recommendations and risk levels for human review.
Actuators adjust dosing, results feed back to the agent to form an adaptive closed loop.
Decisions outside critical thresholds require human confirmation; all actions are audit-logged per regulation.
Side by side
| Dimension | Operator experience alone | Fixed PID control | Agent-driven with human oversight |
|---|---|---|---|
| Basis for the decision | A senior operator's experience and shift habits — it varies with the person | A pre-tuned parameter curve, bound to the conditions it was tuned for | Second-level quality, flow and dosing data, with threshold models and LLM reasoning producing the recommendation together |
| Keeping up when conditions shift | Someone has to notice on a round and then adjust; the lag equals the inspection interval | Parameters stay fixed; going outside the tuned range means an engineer must retune | Actuation results flow back through the sensors, forming an adaptive loop that keeps correcting |
| Who owns out-of-bounds situations | Entirely personal judgement, with blurry accountability | The controller does not assess risk; anomalies still depend on a human noticing | Anything outside the critical thresholds requires human confirmation; the agent only issues a recommendation and a risk level |
| Traceability | Paper logs and verbal handover — hard to reconstruct afterwards | Records the control values but not the reasoning behind them | Recommendation, confirmation and actuation are all audit-logged for industry compliance reviews |
How the agent and human supervision collaborate
source
Sensor array
agent
Agent layer
human
Operator
output
Dosing actuator
Flow
Verifiable facts
The design is a four-stage closed loop: sensor array → agent decision layer → operator approval → dosing actuator, with the result flowing back through the sensors.
The agent never acts alone: decisions outside the critical thresholds require operator confirmation, and every action is audit-logged to meet industry compliance requirements.
Where it fits