Every model, prompt and policy that can touch a plant carries a version, and the shift record names the versions that produced it.
Every release that can touch a plant, written down.
Nothing has shipped to a plant yet. This is the release line the first design partners will see, in the order it is planned, plus the policy that governs how any release reaches a plant.
Illustrative scenario: one municipal activated-sludge wastewater plant with a membrane train, planning the next 24 hours after a rain forecast. Not a measured run.
The planned release line
Status is honest: everything is queued. Order follows the first-release plan.
-
v0.1Connectors and shadow mode QUEUED
Connectors to the wedge process's SCADA, dosing, aeration or membrane controls, online analyzers, lab system and utility MES. Read-only. Predictions logged against lab, analyzer and SCADA outcomes; nothing written.
-
v0.2Review console and audit log QUEUED
The console where process engineers see predictions beside lab and analyzer results, correct them, and feed the models. The immutable, assurance-grade compliance and water-quality audit log from day one.
-
v0.3Assist mode on the wedge QUEUED
Dose, DO or membrane recommendations with predicted quality, energy and permit impact and a citation, approved by an engineer. Accuracy and ROI measured against the baseline.
-
v0.4Permia gating and reporting QUEUED
Continuous conformance evaluation, excursion-risk forecasting and drafted regulatory reports for the design-partner permit set; compliance gating ready for the first bounded loop.
-
v0.5Bounded control, one loop QUEUED
Bounded write-back on one declared low-risk loop inside rate and magnitude limits, with policy evaluation, rollback and the record for every write.
-
v0.6Simflo surrogate twin QUEUED
A calibrated surrogate twin for one process train assimilating live telemetry, with pre-shift what-ifs and candidate-plan evaluation; synthetic rare events to follow.
How a release reaches a plant
Golden datasets and model-as-judge evaluations gate every model and prompt change in CI before it is promoted.
A release reaches a plant in shadow first, then assist, then bounded, per loop, with the design partner's sign-off at each gate.
Rollback by version, per loop, at any time, with the reason recorded.
The reference scenario each release is tested against
Every Hydrira run is an ordered, inspectable sequence. This is run_scn_0418, an illustrative scenario: one municipal activated-sludge wastewater plant with a membrane train, planning the next 24 hours after a rain forecast. Ten steps, nine automatic, one human approval on the blower setpoint move, every action written to the shift record.
-
01
Read influent and forecast
SUCCEEDED
1.2 s
Intake-and-Screen · Flow, turbidity, ammonia and alkalinity at the headworks, and a 24-hour rain forecast that raises the expected influent by mid-afternoon.
-
02
Rehearse the next 24 hours
SUCCEEDED
4 min 10 s
Plant orchestrator · Simflo evaluates candidate dose, DO and flux plans over the 24-hour horizon and keeps the plans that hold permit margin through the surge.
-
03
Fuse cameras and analyzers
SUCCEEDED
180 ms
Quality-and-Anomaly · Clarifier blanket, foam and colour from the basin cameras fused with turbidity, pH, nutrient and residual analyzers into one calibrated plant state.
-
04
Compute the dose curve
SUCCEEDED
2.4 s
Dose-and-Coagulate · Coagulant and polymer dose ramp for the surge window under turbidity, phosphorus, sludge and chemical-cost constraints, cited to the plant's dosing SOP.
-
05
Trim the coagulant setpoint
SUCCEEDED
0.6 s
Dose-and-Coagulate · A bounded setpoint move inside the declared rate and magnitude limits for the dosing pump, with the rollback value recorded first.
-
06
Minimum-energy DO schedule
SUCCEEDED
3.1 s
Aerate-and-Bio · Blower and DO setpoint schedule across the basins that meets the nitrogen and phosphorus targets at the lowest kilowatt-hours the biological state allows.
-
07
Blower move approved
APPROVAL
6 min 02 s
Plant orchestrator · The DO setpoint change is above the plant's threshold for this loop. The process engineer on shift reviews the predicted nutrient margin and approves.
-
08
Flux hold and backwash timing
SUCCEEDED
1.8 s
Membrane-and-Disinfect · Flux held through the surge and the next backwash moved earlier, inside the integrity envelope; the disinfection residual minimum is untouched.
-
09
Permit conformance check
SUCCEEDED
0.9 s
Compliance-and-Energy · Every planned setpoint checked against the discharge permit, the nutrient limits and the residual minimum, with excursion risk forecast for the horizon.
-
10
Write the shift record
SUCCEEDED
0.3 s
Quality-and-Conformance · Inputs, model versions, predictions, limits, the approval and every write appended to the immutable shift record the plant owns.
What the tool log will show per release
Every tool invocation, argument and result is written to an immutable, human-readable log, and every reasoning step is expandable. Nothing about a run is hidden from the plant. The log below is the scenario run, not a measured one.
- 05:40:02scada.read_influent({ plant: "WWTP-A", tags: ["FLOW", "NTU", "NH4", "ALK"] })
- 05:40:03ok flow 412 L/s · turbidity 118 NTU · ammonia 31 mg/L · alkalinity 184 mg/L
- 05:40:03weather.read_forecast({ horizon_h: 24 })
- 05:40:04ok rain from 14:00 · expected influent +38% by 17:00
- 05:40:04twin.simulate({ horizon_h: 24, candidates: 240, objective: ["permit_margin", "kwh", "chemical"] })
- 05:44:14ok 240 plans evaluated · 31 hold permit margin through the surge · best plan selected
- 05:44:15vision.fuse_state({ cameras: ["CLAR-1", "CLAR-2", "AER-3"], analyzers: "all" })
- 05:44:15ok blanket 1.1 m · foam none · effluent turbidity 1.8 NTU · residual 0.9 mg/L
- 05:44:16cuopt.solve_dose({ window_h: 12, constraints: ["NTU", "P", "sludge", "cost"] })
- 05:44:18ok coagulant ramp 38 → 46 mg/L from 13:30 · polymer unchanged · cite sop:DOSE-07 §3
- 05:44:19policy.evaluate({ loop: "dose", stage: "bounded", tag: "DOS1.SP" })
- 05:44:19ok permitted · max_step 4 mg/L per 15 min · rollback 38 mg/L
- 05:44:19plc.write_setpoint({ tag: "DOS1.SP", value: 42, unit: "mg/L" })
- 05:44:20ok written · ack from PLC · audit id 0418-05
- 05:44:20cuopt.solve_blowers({ basins: 4, constraints: ["NH4", "NO3", "P", "DO_min"] })
- 05:44:23ok DO 2.0 → 1.6 mg/L basins 1–3 from 06:00 · predicted −9% kWh · nutrient margin held
- 05:44:24policy.evaluate({ loop: "aeration", stage: "assist", tag: "BLW.DO_SP" })
- 05:44:24above threshold for this loop · routed to approve.human · engineer on shift notified
- 05:50:26approved by the process engineer on shift · reason: nutrient margin 0.6 mg/L above limit
- 05:50:27membrane.schedule_backwash({ train: "MBR-2", surge_at: "14:00" })
- 05:50:29ok flux held 24 LMH · backwash moved 16:00 → 13:15 · residual minimum untouched
- 05:50:29permit.evaluate({ permit: "NPDES-WWTP-A", horizon_h: 24 })
- 05:50:30ok all parameters inside limit · excursion risk 24 h: low · cite permit §II.A
- 05:50:30audit.commit({ run: "run_scn_0418", steps: 10, approvals: 1 })
- 05:50:30ok shift record written · immutable · exportable
Reasoning trace
Goal, plan, thought, action, observation, exactly as the orchestrator would record it.
GoalHold effluent turbidity, nutrients and disinfection residual i…
Hold effluent turbidity, nutrients and disinfection residual inside the discharge permit for the next 24 hours while taking aeration energy and coagulant use to the minimum the permit margin allows.
PlanRead influent and the forecast; rehearse the horizon on Simflo…
Read influent and the forecast; rehearse the horizon on Simflo; fuse the plant state; compute dose and DO schedules; write the dose inside bounds; route the DO move for approval; hold flux and move the backwash; check the permit; commit the record.
ThoughtRain arrives mid-afternoon and the influent will rise by about…
Rain arrives mid-afternoon and the influent will rise by about a third. Dosing has to ramp before the surge, not after, and the DO schedule has to keep nutrient margin through it.
Actioncuopt.solve_dose over a 12-hour window under turbidity, phosph…
cuopt.solve_dose over a 12-hour window under turbidity, phosphorus, sludge and cost constraints; plc.write_setpoint on the coagulant pump inside the declared rate and magnitude limits.
ObservationThe dose ramp holds predicted effluent turbidity under 2 NTU t…
The dose ramp holds predicted effluent turbidity under 2 NTU through the surge. The DO schedule saves energy but is above the threshold set for the aeration loop.
ThoughtEnergy never outranks permit margin. The DO change goes to the…
Energy never outranks permit margin. The DO change goes to the engineer with the predicted nutrient margin attached, and nothing is written until it is approved.
Actionapprove.request to the process engineer on shift; membrane.sch…
approve.request to the process engineer on shift; membrane.schedule_backwash to hold flux and move the next backwash ahead of the surge; permit.evaluate across the horizon.
OutcomeTen steps, one human approval, all parameters inside limit on …
Ten steps, one human approval, all parameters inside limit on the forecast, and the shift record written. Illustrative scenario; nothing here is a measured plant result.
What comes after the first release
From the roadmap, by horizon. Every item is planned, not done.
| Horizon | Planned |
|---|---|
| 0–6 months | Two to four design-partner utilities or industrial sites on the wedge process; MVP: connectors, agent loop, review console, assurance-grade compliance and water-quality audit log; accuracy and energy, chemical and compliance ROI proven in shadow then assist; SOC 2 Type I and critical-infrastructure cyber review; IP-safe and edge deployment |
| 6–12 months | Design partners convert to paid; graduated autonomy on low-risk dosing, aeration and membrane loops; Simflo hit-compliance-before-shift; second process per plant; SOC 2 Type II; SSO and RBAC; first utility and industrial logos; compliance and cyber assurance packages |
| 1–3 years | The full intake, dosing, clarification and filtration, membrane and disinfection, aeration and biological loop; single plant to multi-plant utility to national and industrial fleets; water-equipment, SCADA vendor and engineering-firm channels; international expansion; benchmark data products |
| 3–10 years | The autonomous system of action for water and wastewater treatment, with adjacent scopes through sludge, reuse and desalination and a platform and API ecosystem |
Four stages, set per loop
A plant does not go from manual to bounded control in one step. Hydrira makes the stage explicit per loop, auditable and reversible at any time, and some decisions never leave the person.
| Stage | What the agent does | What the person does | When |
|---|---|---|---|
| Shadow | Predicts water quality, fouling, excursion risk and energy; writes nothing | Runs the plant as today; compares predictions to lab, analyzer and SCADA outcomes | First phase with every design partner |
| Assist | Recommends dose, DO, flux and cleaning moves with predicted quality, energy and permit impact | Approves, corrects or rejects each move; corrections train the models | After accuracy is proven against the plant's own baseline |
| Bounded | Writes to an allow-list of low-risk loops inside declared rate and magnitude limits | Sets the limits, watches the record, pulls the loop back at any time | Per loop, after the assist-mode gate and compliance sign-off |
| Exceptions | Escalates anything above the threshold, every permit-adjacent move and every release | Decides; the operator of record is never removed from compliance decisions | Always |
Physical AI, at the plant edge
A treatment plant runs around the clock on variable influent, and many utilities keep plant data on site, so perception and control inference run at the plant edge. Hydrira keeps training, simulation and optimization in the cloud, in a private cloud or on-prem. Items marked planned are on the roadmap, not running.
Plant-edge inference
TensorRT-optimized time-series and vision models at the plant edge; DeepStream pipelines for clarifier, basin and filter cameras; Holoscan fusing camera streams with turbidity, pH, chlorine and nutrient analyzers at 1 to 10 Hz. Jetson Thor is planned.
Model serving
Water-quality forecasting, anomaly detection, control recommendation and compliance retrieval served with per-tenant routing, model versioning and an audit id on every call. Self-hosted for utility on-prem and edge deployment. Planned.
Training and fine-tuning
Time-series transformers, vision models and NeMo-tuned process and compliance reasoning trained on plant telemetry, operator corrections, images, alarms, lab results, permits and SOPs. Design target: weekly fine-tuning per design-partner cohort, then monthly fleet refreshes. Planned.
The as-run twin
Simflo runs plant hydraulics, clarifier and filtration behaviour, membrane fouling, oxygen transfer and biological response as a GPU simulation calibrated to the live plant, with Modulus surrogates for fast evaluation. Planned.
Synthetic rare events
Wet-weather surges, sensor drift, foaming, fouling, filtration breakthrough and biological upsets generated as scenario variants to stress-test autonomy and train the perception and control agents. Planned.
Optimization
Blower and pump scheduling, chemical dosing windows, membrane cleaning, backwash timing and fleet maintenance routes solved under energy tariff, labour, asset and permit constraints.
GPU analytics
ETL over high-volume SCADA tags, lab results, membrane data, alarms and image metadata to build features and golden evaluation sets. Planned.
Robotics
Supervised robotic sampling routes, asset-inspection rovers and materials handling across intake, treatment and disinfection, where plant policy allows robots. Planned.
An agent that can move a setpoint needs a leash
Hydrira writes to control equipment in a plant that keeps a community's water safe. Every capability is scoped, every write is policy-checked, every output is cited, and every action is written to an append-only audit log the plant owns.
- Bounded action space. Each agent can only write to an explicit tag allow-list, inside per-tag rate and magnitude limits, at the autonomy stage the plant has set.
- Permit limits are read, never written. Permit-limit setpoints, disinfection residual minimums, integrity limits and discharge release decisions stay with the permit and the operator of record.
- Policy engine before every write. Autonomy stage, interlock state, permit margin, excursion forecast and operator presence are evaluated before a setpoint moves.
- Grounded and cited. Every recommendation carries a citation to the permit, SOP or regulation it relied on. Retrieval is tenant-isolated and permission-aware.
- Immutable audit log. Append-only, exportable and retained on the plant's own storage, aligned to Safe Drinking Water Act, discharge-permit and EU Water Framework record-keeping.
- Hard fallback. If the edge node, the network or a model stops, the plant keeps running on its existing SCADA, PLC and DCS control and interlocks.
- Tenant isolation and critical-infrastructure posture. Telemetry, permits, SOPs and models never cross a customer boundary. On-prem and edge-only deployment available for utilities.
Assurance posture
| Item | Scope | Status |
|---|---|---|
| SOC 2 Type I + critical-infrastructure cyber review | Cloud control plane, edge | QUEUED Roadmap, first 6 months |
| SOC 2 Type II | Cloud control plane | QUEUED Roadmap, 6 to 12 months |
| SDWA · discharge permits · EU WFD · nutrient and PFAS rules | Grounding and record-keeping | RUNNING Design basis for retrieval and the audit log |
| Per-tenant isolation | Data, memory, models | RUNNING Design rule |
| On-prem or edge-only deployment | Utilities, critical infrastructure | RUNNING Design option |
Hydrira supports your compliance records. It is not a certification, it does not replace your permit and it never removes the operator of record from a compliance decision.
It speaks plant, not cloud
Hydrira reads and writes through the systems already in the plant. No parallel historian, no new HMI, no rip-and-replace. Connectors are built against each design partner's stack first; long-tail integrations are out of scope for the first release.
Plant SCADA, historians, PLC and DCS interfaces
Tag reads at 1 to 10 Hz, guarded setpoint writes, interlocks untouched
Coagulant, polymer, pH-adjust and disinfectant dosing controllers
Dose reads, bounded setpoint writes inside declared limits
Blower controls, DO and ammonia-based aeration control
DO and blower reads, bounded setpoint writes after approval
Membrane and RO skid PLCs, TMP, flux, conductivity, integrity monitors
Reads; flux and backwash writes inside the integrity envelope
Turbidity, pH, chlorine residual, nutrient, TOC and flow analyzers
Reads; drift and fault flags
Lab information systems, jar-test and grab-sample records
Reads for labels and calibration
Fixed cameras on clarifiers, basins, filters and membrane rooms; inspection robots
Streams fused at the edge; supervised routes
Utility MES, maintenance systems, permit and reporting systems, weather
Reads; shift record and report drafts written back
Straight answers
The questions operations managers, process engineers and compliance leads ask in the first meeting.
Only at the autonomy stage your plant has set, and only to an explicit allow-list of low-risk loops inside documented limits. Every design partner starts in shadow mode, where Hydrira predicts and writes nothing. Assist mode follows, where a process engineer approves each move. Bounded write-back comes last, per loop, after the accuracy gate and compliance sign-off, and can be pulled back at any time.
No. Permit-limit setpoints, disinfection residual minimums, integrity limits and discharge release decisions are read and checked, never written. Permia evaluates every planned move against the permit set and blocks anything that would thin a margin. The operator of record stays the operator of record.
The plant keeps running on its existing SCADA, PLC and DCS control and interlocks. Hydrira is supervisory, not primary. Loss of Hydrira is loss of advice, and the gap is recorded in the shift record.
Shadow mode starts with the SCADA tags, analyzer streams and lab records you already have. Accuracy is measured against your own lab and analyzer baseline, and the first-release plan converts to assist mode only when the number is proven. The design targets assume telemetry you already collect, not new instrumentation.
Telemetry, permits, SOPs and models are tenant-scoped, and an on-prem or edge-only deployment is available for utilities. Cross-plant learning is federated and IP-isolated, planned and opt-in: patterns for similar basins, membranes and influent regimes move, raw data does not.
Not yet, and the site does not claim it. SOC 2 Type I with a critical-infrastructure cyber review is on the roadmap for the first six months, Type II for the six to twelve month window. Safe Drinking Water Act, discharge permit, EU Water Framework and nutrient and PFAS rules are what the retrieval grounds on and what the audit log is aligned to; that is a design basis, not a certification.
The one with the clearest number. Aeration energy on an activated-sludge plant, coagulant cost on a dosing train or membrane run length on an RO or MBR train are the usual wedges, because kilowatt-hours, chemical invoices and cleaning counts are already measured. Turbix is the sensing wedge where visibility is the first problem.
Two to four utilities or industrial sites instrument one wedge process in shadow mode and baseline energy, chemicals and compliance from their own records. Assist mode proves accuracy and ROI with an engineer approving each move. Graduated autonomy follows on low-risk loops. Expansion adds adjacent processes, then fleet sites. Gates, not dates.
Follow the release line
Design partners see each release before it reaches their plant. Ask to be one.