Developers

Autonomy policy is declared. Review it like code.

There is no public SDK yet. What exists is the integration surface design partners work with: connectors, an autonomy policy per loop, cited recommendations, versioned models and an audit record for every action. This page describes that surface honestly.

Talk to the team What the docs cover

ConnectorsPolicy per loopAudit records

run_scn_01 · agent graph · scenario 10 nodes · 1 approval gate
01 · orchestrator order.receive ✓ SUCCEEDED 02 · orchestrator twin.simulate ✓ SUCCEEDED 03 · brew_blend brew.blend ✓ SUCCEEDED 04 · ferment_yeast ferment.control ✓ SUCCEEDED 05 · filter_carbonate carbonate.dose ✓ SUCCEEDED 06 · fill_seal fill.seal ✓ SUCCEEDED 07 · defect_inspect inspect.line ✓ SUCCEEDED 08 · pasteurize_cip pasteurize.verify ✓ SUCCEEDED 09 · orchestrator approve.human ◆ APPROVAL 10 · quality_conf release.genealogy ✓ SUCCEEDED

Illustrative scenario: Plant run · one lager batch from brewhouse to can · brewhouse, cellar, filler, tunnel pasteurizer. Not a measured run.

Developers

Every action is a record

There is no SDK to download yet. What design partners get is the integration surface: connectors, an autonomy policy declared per loop, and an audit record for every reading, recommendation, write and approval. This is the shape of one record, illustrative.

What the surface gives you

  • Connectors to PLC and SCADA, inline analyzers, lab systems, MES, cameras and meters, with declared allow-lists and limits
  • An autonomy policy per loop: stage, tags, rate and magnitude limits, named approvers
  • Citations on every recommendation to the recipe, spec or procedure it relied on
  • Model versions named in every record, with rollback by version
  • Export of the audit log and batch genealogy to your quality system

What the docs cover Developer overview

audit record · illustrative 
// one write, as the audit log would hold it
{
  "run": "run_scn_01",
  "step": "fill.seal",
  "agent": "agent.fill_seal",
  "stage": "bounded",
  "action": { "tool": "plc.write_setpoint", "tag": "FIL1.VALVE_14" },
  "limits": { "max_step": "declared per tag", "per": "declared per tag" },
  "evidence": ["vision.stream", "filler.read_telemetry"],
  "citation": "spec:CAN-330 fill-level v3 §2",
  "policy": "permitted",
  "models": { "fill_level": "versioned", "policy": "versioned" },
  "human": null,
  "outcome": "on target level · inside approved window"
}
Capabilities

What the surface gives you

Connectors

Declared read and write scopes for brewhouse, cellar, skid, filler, pasteurizer and inspection PLCs, analyzers, lab, MES, cameras and meters.

Autonomy policy

Stage, tag allow-list, rate and magnitude limits and named approvers per loop, versioned like any other change.

Cited recommendations

Every recommendation names the recipe, spec or procedure it relied on; retrieval is tenant-isolated and permission-aware.

Model router

The best or cheapest model serves each step; high-volume line vision moves to fine-tuned open models, reasoning steps use frontier models.

Evaluation gates

Golden datasets and model-as-judge evaluations gate model and prompt changes in CI.

Export

Audit log and genealogy export to the plant's quality system.

Run timeline

What a run record contains

Every Brewira run is an ordered, inspectable sequence. This is run_scn_01, an illustrative scenario: Plant run · one lager batch from brewhouse to can · brewhouse, cellar, filler, tunnel pasteurizer. Scenario complete · batch inside spec · one human approval at the pasteurizer.

run_scn_01 scenario · not measured SUCCEEDED
  1. 01 order.receive SUCCEEDED 0.8 s

    Plant Orchestrator · Pulled the recipe, the package spec, the HACCP plan with its pasteurization critical control point, the yeast lot record and the standing CIP window from the beverage MES and the lab system; locked the target envelope for the run.

  2. 02 twin.simulate SUCCEEDED 38 s

    Plant Orchestrator · Simulated 24 candidate brew, fermentation and carbonation plans on the Cuvera as-produced twin; ranked them on gravity margin, predicted finish time, CO2 target, PU margin, takt and CO2 and energy use, and discarded every candidate that thinned the pasteurization margin.

  3. 03 brew.blend SUCCEEDED 2 m 40 s

    Brew-and-Blend · Held wort gravity to recipe from the inline gravity analyzer and the brewhouse PLC, trimming the blend and utilization setpoints inside the recipe window rather than correcting after a spot lab check.

  4. 04 ferment.control SUCCEEDED 5 d 06 h

    Ferment-and-Yeast · Predicted the attenuation curve and finish time for the tank from inline gravity, dissolved oxygen and CO2, pH and temperature; held the temperature ramp and timed the crash inside the brewmaster-approved window; no stuck or runaway fermentation was flagged.

  5. 05 carbonate.dose SUCCEEDED 3 m 15 s

    Filter-and-Carbonate · Fused inline CO2, dissolved oxygen, turbidity, temperature and pressure; held carbonation to the package target with the minimum CO2 and timed the filter backwash from the predicted loading instead of a fixed schedule.

  6. 06 fill.seal SUCCEEDED continuous

    Fill-and-Seal · Adjusted fill valves toward the target level inside the approved window using drift feedback from inspection, and ran seam checks on the seamer; the filler corrected upstream instead of rejecting downstream.

  7. 07 inspect.line SUCCEEDED continuous

    Defect-and-Inspect · 16 camera streams at the line edge: fill level, seam and cap, label and date code, foreign body and container checks on every can; three low-fill rejects were attributed to one filler valve and Fill-and-Seal corrected it.

  8. 08 pasteurize.verify SUCCEEDED continuous

    Pasteurize-and-CIP · Held the tunnel pasteurizer on the validated PU profile at the lowest zone energy inside the envelope; every critical control point reading was written to the audit log against the HACCP plan and the CIP release checks were armed.

  9. 09 approve.human APPROVAL 1 m 52 s

    Plant Orchestrator · The lowest-energy candidate wanted to move the pasteurizer CIP into a window that overlapped the next brand's start. Held for the shift quality manager, who kept the validated PU profile untouched and approved the CIP move with a tighter release check.

  10. 10 release.genealogy SUCCEEDED 1 m 05 s

    Quality-and-Conformance · Batch released: gravity and attenuation curves, yeast lot, CO2 and clarity, fill and seal outcomes, inspection rejects, PU readings, CIP release and every human decision written to the batch genealogy and checked against spec.

Tool calls

Streaming a run

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.

tool-call stream · run_scn_01
  1. 04:50:02mes.read_order({ batch: "L-0931", brand: "LAGER-33", package: "CAN-330" })
  2. 04:50:02ok recipe locked · HACCP plan v6 · CCP: tunnel PU 15–20
  3. 04:50:03lab.read_results({ yeast_lot: "Y-2207", test: "viability" })
  4. 04:50:03ok viability inside window · lot accepted for pitching
  5. 04:50:04twin.simulate({ candidates: 24, horizon: "batch" })
  6. 04:50:42ok best plan #7 · PU margin held · lowest CO2 and energy per hl
  7. 04:50:43policy.evaluate({ plan: 7, stage: "bounded" })
  8. 04:50:43ok 7 writes permitted · 1 write requires human approval
  9. 04:50:44plc.write_setpoint({ tag: "BH1.BLEND_RATIO", basis: "gravity+recipe" })
  10. 04:53:24ok wort gravity on recipe · utilization held
  11. 05:10:00ferment.predict_curve({ tank: "FV-12", lot: "Y-2207" })
  12. 05:10:00ok attenuation curve predicted · finish day 5 · crash timed
  13. 05:10:01plc.write_setpoint({ tag: "FV12.TEMP_RAMP", window: "brewmaster_approved" })
  14. 05:10:01ok ramp applied · no stuck or runaway ferment flagged
  15. 06:02:10plc.write_setpoint({ tag: "CARB1.CO2_DOSE", basis: "co2+do+turbidity" })
  16. 06:05:25ok CO2 on package target · clarity inside spec · backwash timed
  17. 06:05:26vision.stream({ cameras: 16, models: ["fill_level", "seam_cap", "label_ocr", "foreign_body"] })
  18. 06:31:48reject: low fill · can 20417 · attributed to filler valve 14 drift
  19. 06:31:49defect.attribute({ event: "low_fill", cause: "filler.valve_14" })
  20. 06:31:49ok fill_seal notified · valve 14 trimmed inside window · two more rejects, then clean
  21. 06:31:50plc.verify_pu({ tag: "TUN1.ZONES", profile: "validated" })
  22. 06:31:50ok PU inside profile · readings → audit log
  23. 06:31:52approval.request({ action: "cip.window", reason: "overlaps next brand start" })
  24. 06:33:44ok approved by the shift quality manager · PU profile untouched · tighter release check
  25. 09:12:30mes.release_batch({ batch: "L-0931", genealogy: "full" })
  26. 09:13:35ok batch released · scenario complete

Reasoning trace

Goal, plan, thought, action, observation, exactly as the orchestrator would record it.

GoalBrew batch L-0931 to recipe gravity, ferment it to target atte…

Brew batch L-0931 to recipe gravity, ferment it to target attenuation, carbonate and fill it to spec at line speed, pasteurize it on the validated PU profile and release it with full genealogy.

PlanSimulate the whole batch on the twin first. Lead with gravity …

Simulate the whole batch on the twin first. Lead with gravity so the cellar gets the right wort and the filler gets the right CO2. Verify the pasteurization critical control point continuously. Let the filler chase target level only inside the approved window; hold anything beyond the envelope for a human.

ThoughtThe standing practice over-fills to avoid low fills and over-g…

The standing practice over-fills to avoid low fills and over-gasses to avoid flat product. With continuous fill-level estimation and inline CO2 the filler and the carbonator can sit at target. The saving is giveaway, not safety margin: the pasteurization CCP is never a candidate for optimization.

Actiontwin.simulate(candidates=24) → plan #7: gravity on recipe, fin…

twin.simulate(candidates=24) → plan #7: gravity on recipe, finish day 5, CO2 at package target, PU margin unchanged, lowest predicted CO2 and energy per hectoliter.

ObservationPlan #7 wants the pasteurizer CIP in a window that overlaps th…

Plan #7 wants the pasteurizer CIP in a window that overlaps the next brand's start. The policy engine routes that single move to the shift quality manager and permits the other seven writes.

ActionExecute plan #7 at the bounded stage: brewhouse blend, tank ra…

Execute plan #7 at the bounded stage: brewhouse blend, tank ramp, CO2 dosing and fill-valve writes proceed; PU is verified continuously; the CIP move waits.

ObservationApproval granted at 06:33:44 with the PU profile untouched and…

Approval granted at 06:33:44 with the PU profile untouched and a tighter CIP release check. Three low-fill rejects were attributed to filler valve 14 and corrected by Fill-and-Seal inside the approved window.

OutcomeScenario complete. The batch is released with gravity and atte…

Scenario complete. The batch is released with gravity and attenuation curves, yeast lot, CO2 and clarity, fill and seal outcomes, PU readings, CIP release and the human decision in the genealogy. No measured plant result is claimed; this is the run the platform is designed to execute.

Workflow

From idea to a line, safely

  1. Declare. Tools, tags and limits for the agent, and the autonomy stage for the loop.
  2. Shadow. Run against live plant data; compare predictions to lab, QA and PLC outcomes; nothing is written.
  3. Evaluate. Golden datasets and model-as-judge checks in CI before a model or prompt version is promoted.
  4. Rehearse. High-impact moves are simulated on Cuvera where a twin format exists.
  5. Assist, then bounded. Writes only after the accuracy and food-safety gates, inside declared limits.
  6. Record and retrain. Every action and correction is logged; corrections become training events.
Autonomy

Four stages, set per loop

A plant does not go from manual to bounded control in one step. Brewira makes the stage explicit per loop, auditable and reversible at any time, and some decisions never leave the person.

Autonomy stages and the human role at each
StageWhat the agent doesWhat the person doesWhen
ShadowPredicts gravity, attenuation, carbonation, fill and seal faults, contamination and takt; writes nothingRuns the plant as today; compares predictions to lab, QA and PLC outcomesFirst phase with every design partner
AssistRecommends bounded brew, fermentation, carbonation, fill and seal, pasteurization and CIP moves with the evidenceA brewmaster, process engineer or quality manager approves or rejects each oneAfter accuracy is proven against the baseline
BoundedWrites low-risk filling, carbonation and inspection setpoints inside documented recipe, quality, food-safety, equipment and rollback limitsApproves anything above the limits; reviews the shift recordAfter the food-safety and ROI gates agreed up front
ExceptionsEscalates anything outside its envelope, CCP-adjacent moves and release decisionsAlways decides these; the agent never doesPermanent
Multi-agent handoff

Agents negotiate, they do not collide

Two agents will want the same actuator. The orchestrator arbitrates on the run goal, with food safety above everything, and the handoff is logged like any other step.

Contested move: tunnel pasteurizer belt speed

  1. 01yield_takt.request(belt +5%) SUCCEEDED0.2 s

    Yield-and-Takt wants more belt speed through the tunnel pasteurizer so the filler is not held back and the shift meets takt.

  2. 02pasteurize_cip.request(belt hold) SUCCEEDED0.2 s

    Pasteurize-and-CIP requires the belt speed that keeps every container inside the validated PU profile. That is the HACCP critical control point.

  3. 03orchestrator.arbitrate SUCCEEDED0.4 s

    Food safety outranks takt under every run goal. The critical control point wins the actuator; the request for more speed is logged with its score and the reason it lost.

  4. 04yield_takt.handoff(resequenced) SUCCEEDED90 s

    Yield-and-Takt recovers the schedule by pulling the next changeover earlier and shifting the CIP window instead. Takt is met without touching the CCP.

Arbitration rules

  • Run goal first. Every request is scored against the declared goal, not against the requesting agent's local objective.
  • Food safety outranks everything. A critical control point is never an optimization variable. An under-pasteurized or contaminated batch is a recall; a takt miss is a schedule.
  • Time-boxed ownership. An agent holds a contested actuator for a bounded window, then must re-justify.
  • Everything is logged. The losing request, the score and the reason all appear in the run record.
  • Deadlocks escalate to a human rather than resolving by timeout.

See the agent roster

Accelerated computing

Physical AI, at the plant edge

Lines fill tens of thousands of containers an hour, so perception has to be local and fast. Brewira runs GPU inference at the line edge and keeps training, simulation and optimization in the cloud, in a private cloud or on-prem. Items marked planned are on the roadmap, not running.

Jetson Orin / Thor · DeepStream · TensorRT · Holoscan

Edge perception

8 to 24 camera feeds per line through DeepStream pipelines and TensorRT detectors, segmentation, OCR and fill-level models; Holoscan synchronizes cameras with gravity, CO2 and clarity analyzers and filler and pasteurizer telemetry. Design target: sub-100 ms reject decisions.

Triton · NIM

Model serving

Vision, time-series, anomaly, process and food-safety reasoning models behind one orchestrator, with version routing by brand, package and line. Design target: 20 to 60 model versions per enterprise tenant.

DGX / HGX · NeMo

Training (planned)

Line vision, time-series forecasters, anomaly models and NeMo-based beverage reasoning fine-tuned from labeled batch outcomes, brewmaster corrections, lab results and release decisions. Design target: weekly retraining for wedge models, monthly for validated process models.

Omniverse · Modulus · OVX

Digital twin (planned)

The Cuvera as-produced line twin: CUDA-accelerated fermentation, carbonation, mixing and pasteurization heat simulation run before the batch, for new SKU setup and batch planning when a trial is expensive.

Cosmos · Omniverse Replicator

Synthetic data (planned)

Rare leaker, low-fill, overfill, cap-skew, seam, label-shift, condensation, glare and foreign-body scenarios generated with domain randomization and validated against real plant events before use.

cuOpt

Optimization

Brew and fill schedules, changeovers, CIP cycles, water and energy allocation and takt balancing under line, tank, product and sanitation constraints, recomputed on the GPU when the plant is disrupted.

RAPIDS

Data engineering

Image, time-series and batch-genealogy ETL for feature generation, quality correlation and cross-plant benchmarks on the GPU.

Isaac · Jetson

Robotic handling

Pack, case and palletizing handling, with perception and motion policies built and validated in simulation before they move a pallet.

Integrations

It speaks plant, not cloud

Brewira reads and writes through the systems already on the floor. 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.

Brewhouse and fermentation

PLC and SCADA of brewhouse, blending and syrup systems, fermentation tanks and yeast handling

Setpoint reads, guarded writes, interlocks untouched

Filtration and carbonation

Filtration and carbonation skids and their inline CO2, dissolved-oxygen and turbidity analyzers

Continuous carbonation and clarity state

Filler, seamer, capper

High-speed filler, seamer and capper controls

Telemetry in; fill-valve and seam settings inside approved windows

Pasteurizer and CIP

Tunnel and flash pasteurizers, CIP skids and hygiene sensors

PU readings verified and logged; CIP release checks

Inspection

Multi-camera stations for fill level, seal and cap, label, foreign body and container defects

Frames, defect events, reject commands

Beverage MES and lab

Orders, recipes, product specs, batch genealogy, LIMS results

Targets in, release records out

Energy, water, robots

Plant and asset meters; Isaac-driven pack, case and palletizing cells

Hard constraints for scheduling; bounded handling

Identity (planned)

SSO and role-based access; SOC 2 on the roadmap

Named approvers, scoped roles

See all integrations

Guardrails

An agent that can move a setpoint needs a leash

Brewira writes to production equipment in a food plant. 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.
  • Critical control points are read, never written. Pasteurization units, CIP release checks and batch release decisions stay with the HACCP plan and the people who own it.
  • Policy engine before every write. Autonomy stage, interlock state, product, shift and operator presence are evaluated before a setpoint moves.
  • Grounded and cited. Every recommendation carries a citation to the recipe, spec or procedure 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 HACCP, FSMA, FSSC 22000 and BRCGS record-keeping.
  • Hard fallback. If the edge node, the network or a model stops, the plant keeps running on its existing PLC and SCADA control and interlocks.
  • Tenant and IP isolation. Recipes, supplier data and tolerance models never cross a customer boundary. On-prem deployment available.

Assurance posture

ItemScopeStatus
SOC 2 Type ICloud control plane QUEUED Roadmap, first 6 months
SOC 2 Type IICloud control plane QUEUED Roadmap, 6 to 12 months
HACCP · FSMA · FSSC 22000 · BRCGSGrounding and record-keeping RUNNING Design basis for retrieval and the audit log
Per-tenant isolationData, memory, models RUNNING Design rule
On-prem deploymentRecipe-sensitive producers RUNNING Design option

Brewira supports your food-safety records. It is not a certification and it does not replace your HACCP plan.

Read the security overview

FAQ

Straight answers

The questions plant directors, brewmasters and quality managers 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 Brewira predicts and writes nothing. Assist mode follows, where a brewmaster or process engineer approves each recommendation. Bounded autonomy on filling, carbonation and inspection loops comes after the food-safety and ROI gates agreed up front.

Get started

Build on the run engine

Design partners get the integration surface and the people who built it. Tell us what you want to connect.

Book a line review See pricing