Inline composition and vision
Fat, protein, solids, moisture and pH from NIR, mid-IR and inline sensors, fused with lab calibration; contamination, foreign-body, fill, seal, cap and label vision at line speed.
Features
Everything a plant needs to trust software that acts.
Autonomy in a food plant is a set of features that make action safe: cited reasoning, graduated control, bounded writes, an immutable audit log, batch genealogy and a twin that rehearses the recipe first. This page lists them one by one.
Illustrative scenario: Plant run · skim-milk-powder line + set-yogurt line · shared intake. Not a measured run.
Core
The features that make autonomy safe
Each of these exists because a dairy plant is safety-critical, hygiene-critical and continuous.
Prognostics with lead time
Separator fouling, clarifier bacterial load, dryer stickiness and wall fouling, abnormal acidification and phage events, predicted early enough to act.
Adaptive setpoints and recipes
Closed loops on separation and standardization, culture profiles and drying, inside explicit limits, at the autonomy stage you set.
Recipe before the run
Lactwin ranks candidate recipes on composition margin, pasteurization margin, energy and takt, and discards any that thin a safety margin.
Cited reasoning
Every recommendation carries a citation to the recipe, spec, HACCP plan or CIP procedure it relied on, through tenant-isolated retrieval.
Audit log and genealogy
Every reading, recommendation, write, approval and release in an append-only log, assembled into the batch genealogy for release.
Craft capture
Per-plant yield-and-quality history and per-technologist corrections, versioned and scoped per tenant, so the next batch starts from what the last one taught.
Defect to cause
Each inspection reject is attributed to the upstream agent and state that caused it, closing the loop across the plant.
Yield, energy, water, takt
cuOpt sequences products, tanks, dryer schedules and CIP windows under energy and water constraints and rebalances when the plant is disrupted.
Run timeline
One scenario run, step by step
Every Milkeon run is an ordered, inspectable sequence. This is run_scn_01, an illustrative scenario: Plant run · skim-milk-powder line + set-yogurt line · shared intake. Scenario complete · both streams inside spec · one human approval at the dryer.
run_scn_01
scenario · not measured
SUCCEEDED
-
01
intake.receive
SUCCEEDED
0.9 s
Plant Orchestrator · Pulled both product specs, the HACCP plan with its pasteurization critical control point, the antibiotic screen for the tanker lot and the standing CIP window from MES and the lab system; locked the target envelope for the run.
-
02
twin.simulate
SUCCEEDED
41 s
Plant Orchestrator · Simulated 24 candidate recipes across separation, pasteurization, evaporation and spray-drying on the Lactwin as-processed twin; ranked them on composition margin, pasteurization margin, energy per kilogram of powder and takt, and discarded every candidate that thinned the pasteurization margin.
-
03
separate.standardize
SUCCEEDED
3 m 05 s
Intake-and-Separate · Fused inline NIR and mid-IR with the latest lab calibration, then trimmed the cream-skim split and the standardization blend valve so skim solids and the yogurt-stream fat sat at target with the minimum safe margin rather than the usual over-standardization.
-
04
pasteurize.verify
SUCCEEDED
continuous
Pasteurize-and-Homogenize · Held HTST at or above 72 °C for 15 s through the holding tube with the flow-diversion valve armed; every critical control point reading was written to the audit log against the HACCP plan.
-
05
culture.vat
SUCCEEDED
4 h 20 m
Ferment-and-Culture · Predicted the acidification curve for the culture lot from inline pH and the upstream composition, held the vat temperature profile and timed the hold so pH landed on the set-yogurt target curve; no abnormal acidification was flagged.
-
06
dry.spray
SUCCEEDED
2 h 10 m
Evaporate-and-Dry · Ran the dryer against the twin's moisture and stickiness prediction; stepped inlet temperature and feed solids toward the minimum-energy setpoint while predicted outlet moisture stayed inside spec and fouling risk stayed low.
-
07
inspect.pack
SUCCEEDED
continuous
Defect-and-Inspect · 16 camera streams across the two packing lines at the plant edge: fill, seal, cap and label checks on every unit, foreign-body and contamination models on the powder stream; two rejects were attributed back to a seal-jaw temperature drift and the packing line was notified.
-
08
cip.schedule
SUCCEEDED
6 s
Hygiene-and-CIP · Placed the pasteurizer and dryer CIP into the cheapest energy window, sequenced wash phases against tank availability and set the release checks that must pass before the next product.
-
09
approve.human
APPROVAL
1 m 48 s
Plant Orchestrator · The twin's lowest-energy candidate wanted a dryer inlet temperature above the site's bounded envelope. Held for the shift dairy technologist, who approved the step with a tighter outlet-moisture alarm.
-
10
release.genealogy
SUCCEEDED
1 m 12 s
Quality-and-Conformance · Both batches released: composition, CCP records, culture lot and pH curve, drying setpoints, inspection rejects and every human decision written to the batch genealogy and checked against spec.
Tool calls
The agent shows its work
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.
- 05:10:02mes.read_order({ lot: "T-0417", products: ["SMP", "SET-YOG"] })
- 05:10:02ok specs locked · HACCP plan v7 · CCP: pasteurization 72 °C / 15 s
- 05:10:03lab.read_results({ lot: "T-0417", test: "antibiotic_screen" })
- 05:10:03ok negative · lot accepted for intake
- 05:10:04twin.simulate({ candidates: 24, horizon: "shift" })
- 05:10:45ok best recipe #9 · pasteurization margin held · lowest energy per kg
- 05:10:46policy.evaluate({ recipe: 9, stage: "bounded" })
- 05:10:46ok 7 writes permitted · 1 write requires human approval
- 05:10:47plc.write_setpoint({ tag: "SEP1.BLEND_VALVE", basis: "nir+midir+lab" })
- 05:13:52ok skim solids on target · yogurt-stream fat at minimum safe margin
- 05:13:53plc.verify_ccp({ tag: "PAST1.HOLD_TUBE", min: "72 °C / 15 s" })
- 05:13:53ok CCP verified · divert valve armed · readings → audit log
- 05:14:10dose.culture({ vat: "Y2", lot: "C-2291", profile: "set-yogurt" })
- 05:14:10ok acidification curve predicted · hold timed to target pH
- 05:14:12approval.request({ action: "dryer.inlet", reason: "above bounded envelope" })
- 05:16:00ok approved by the shift dairy technologist · tighter moisture alarm set
- 05:16:01plc.write_setpoint({ tag: "DRY1.INLET_T", mode: "min_energy" })
- 05:16:01ok predicted outlet moisture inside spec · fouling risk low
- 05:16:02vision.stream({ cameras: 16, models: ["fill_seal", "foreign_body"] })
- 05:41:18reject: seal fault · unit 1184 · attributed to seal-jaw temperature drift
- 05:41:19defect.attribute({ event: "seal_fault", cause: "pack_line.jaw_temp" })
- 05:41:19ok packing line notified · second reject at 05:43:02 · drift corrected
- 05:41:20cuopt.schedule({ assets: ["PAST1", "DRY1"], window: "energy_low" })
- 05:41:20ok CIP sequenced · release checks armed
- 09:36:40mes.release_batch({ batches: ["SMP-0417", "YOG-0417"], genealogy: "full" })
- 09:37:52ok both batches released · scenario complete
Reasoning trace
Goal, plan, thought, action, observation, exactly as the orchestrator would record it.
GoalReceive tanker lot T-0417, standardize it once, pasteurize onc…
Receive tanker lot T-0417, standardize it once, pasteurize once, and run a skim-milk-powder stream and a set-yogurt stream to spec with full genealogy.
PlanSimulate the whole recipe on the twin first. Lead with composi…
Simulate the whole recipe on the twin first. Lead with composition so the dryer gets the right solids and the vat gets the right fat. Verify the pasteurization critical control point continuously. Let the dryer chase minimum energy only inside the bounded envelope; hold anything beyond it for a human.
ThoughtThe standing recipe over-standardizes both streams to protect …
The standing recipe over-standardizes both streams to protect against lab-sample lag. With continuous NIR and mid-IR estimates the blend can sit at the minimum safe margin. The saving is composition give-away, not safety margin: the pasteurization CCP is never a candidate for optimization.
Actiontwin.simulate(candidates=24) → recipe #9: minimum safe composi…
twin.simulate(candidates=24) → recipe #9: minimum safe composition margin, pasteurization margin unchanged, lowest predicted energy per kilogram of powder.
ObservationRecipe #9's dryer inlet temperature sits above the site's boun…
Recipe #9's dryer inlet temperature sits above the site's bounded envelope. The policy engine routes that single write to the shift dairy technologist and permits the other seven.
ActionExecute recipe #9 at the bounded stage: separation and standar…
Execute recipe #9 at the bounded stage: separation and standardization writes, CCP verification, culture dosing and the vat profile proceed; the dryer inlet write waits.
ObservationApproval granted at 05:16:00 with a tighter outlet-moisture al…
Approval granted at 05:16:00 with a tighter outlet-moisture alarm. Two seal rejects on the yogurt packing line were attributed to seal-jaw temperature drift and corrected by the packing line.
OutcomeScenario complete. Both batches released with composition, CCP…
Scenario complete. Both batches released with composition, CCP records, culture lot, drying setpoints, inspection events and the human decision in the genealogy. No measured plant result is claimed; this is the run the platform is designed to execute.
In depth
Three features worth the detail
Latency
Decisions at the line, not in the cloud
A reject decision is only useful if it arrives before the unit leaves the station, and a composition estimate is only useful if it arrives before the next blend move. The design targets are 50 to 150 ms and 250 to 1,000 ms respectively, on plant-edge GPUs, with the cloud used for training and simulation.
| Decision | Design target |
|---|---|
| Reject or alert | 50–150 ms p95 |
| Composition estimate | 250–1,000 ms p95 |
| Streams per line | 8–24 |
Evidence
Nothing is recommended without a citation
The orchestrator retrieves over product and composition recipes, processing and culture specs, food-safety standards (HACCP, FSMA, ISO 22000, FSSC, Codex) and quality, genealogy and CIP procedures. Retrieval is multi-tenant isolated and permission-aware, and every output is traceable to a source.
- Citation and grounding checks on every output
- Schema and safety guardrails
- Golden datasets and model-as-judge evaluation in CI
- Immutable, assurance-grade audit log for every action
Rehearsal
The twin discards what the plant should never try
Lactwin simulates the batch as it will actually be processed, across the coupled physics and biology of separation, heat exchange, fermentation, evaporation and spray-drying. Candidates that save energy but thin the pasteurization margin never reach the agents.
- Multi-objective ranking: composition, safety margin, energy, takt
- Safe control envelopes handed to the agents
- Twin-versus-actual reconciliation every batch
- Validated synthetic rare faults for training
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: pasteurizer flow rate
-
01yield_takt.request(flow +6%) SUCCEEDED0.2 s
Yield-and-Takt wants more flow through the pasteurizer so the powder stream starts earlier and the dryer is not left waiting.
-
02pasteurize_homog.request(flow hold) SUCCEEDED0.2 s
Pasteurize-and-Homogenize requires the flow that keeps holding time at or above 15 s at 72 °C. That is the HACCP critical control point.
-
03orchestrator.arbitrate SUCCEEDED0.4 s
Food safety outranks takt under every run goal. The critical control point wins the actuator; the request for more flow is logged with its score and the reason it lost.
-
04yield_takt.handoff(resequenced) SUCCEEDED90 s
Yield-and-Takt recovers the schedule by starting the evaporator on the first standardized skim instead of waiting for the full tank. 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. A pasteurization or seal escape 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.
Autonomy
Four stages, set per loop
A plant does not go from manual to bounded control in one step. Milkeon 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 composition, contamination, moisture and takt; writes nothing | Runs the plant as today; compares predictions to lab results | First phase with every design partner |
| Assist | Recommends separation, pasteurization, culture and drying moves with the evidence | A dairy technologist or process engineer approves or corrects each one | After accuracy is proven against the baseline |
| Bounded | Writes low-risk separation, drying and inspection setpoints inside explicit limits | Approves anything above the limits; reviews the shift record | After the food-safety and ROI gates agreed up front |
| Exceptions | Escalates anything outside its envelope, CCP-adjacent moves and release decisions | Always decides these; the agent never does | Permanent |
How it works
A goal becomes a plan becomes a run
Nothing about a Milkeon run is implicit. The goal is declared, the plan is rehearsed on the twin, each step is typed, and the outcome is written to the batch genealogy.
The orchestrator plans, the agents act
A run starts as a goal, becomes a plan, and executes as a graph of typed steps. The active path is always visible, always logged, always reversible.
RUNNING APPROVAL SUCCEEDED
Autonomy is a dial, not a switch
Shadow, assist, bounded, exceptions. Each plant sets the stage per loop, and anything above the threshold, every critical control point and every release decision waits for a named person.
Design targets
Feature design targets
The numbers Milkeon is engineered against, from the GPU utilization plan.
50–150 ms
p95 latency for reject and alert decisions at the plant edge
250–1,000 ms
p95 latency for inline composition estimates
8–24
camera and inline-analyzer streams per line
10–40
concurrent models served per plant, routed by product, line and sensors
2k–10k
GPU-hours of training across the first design-partner cohort
50k–250k
synthetic rare-fault scenarios, each validated against real plant data
Design targets from the GPU utilization plan. Nothing on this page is a measured plant result; measurements start when the first design-partner line runs in shadow mode.
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
// one write, as the audit log would hold it
{
"run": "run_scn_01",
"step": "separate.standardize",
"agent": "agent.intake_separate",
"stage": "bounded",
"action": { "tool": "plc.write_setpoint", "tag": "SEP1.BLEND_VALVE" },
"limits": { "max_step": "declared per tag", "per": "declared per tag" },
"evidence": ["nir.read_composition", "lab.read_results"],
"citation": "spec:SMP-standardization v4 §2",
"policy": "permitted",
"models": { "composition": "versioned", "policy": "versioned" },
"human": null,
"outcome": "inside spec · minimum safe margin"
}
From the plant floor
Who this is built for
The three people in a dairy plant whose week changes when the process runs itself, and what each of them needs from Milkeon.
Plant or Operations Director
Owns production rate, quality, loss and yield, energy and water, and food safety. Wants a plant that holds spec at higher rate with fewer interventions, and a number that justifies the spend.
Dairy Technologist or Process Engineer
Owns separation and standardization, pasteurization, culture and drying recipes, process windows and yield. Approves and corrects the agents in assist mode; those corrections train the models.
Quality and Food-Safety Manager
Owns HACCP, FSMA, ISO 22000 and FSSC conformance, composition and contamination conformance, and batch genealogy. Needs every action grounded, cited and written to an immutable record.
IT/OT and sustainability leads approve the deployment. There are no customer quotes on this site yet; design partners will speak for themselves when they choose to.
Integrations
It speaks plant, not cloud
Milkeon 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.
PLC and SCADA of separators, pasteurizers, homogenizers, culture vats, evaporators, spray dryers and CIP skids
Setpoint reads, guarded writes, interlocks untouched
NIR and mid-IR composition, inline pH, moisture, temperature and humidity
Continuous composition and process state
LIMS results, calibrations, antibiotic and micro screens
Calibration and release evidence
Orders, recipes, product specs, batch genealogy
Targets in, release records out
Multi-camera stations on filling, sealing, capping, labeling and powder lines
Frames, defect events, reject commands
Plant and asset-level meters
Hard constraints for the dryer and CIP schedule
Isaac-driven packaging, palletizing and handling cells
Perception and motion for packs and pallets
SSO and role-based access; SOC 2 on the roadmap
Named approvers, scoped roles
Guardrails
An agent that can move a setpoint needs a leash
Milkeon 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 time-temperature, flow diversion and 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, ISO 22000 and FSSC 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
| Item | Scope | Status |
|---|---|---|
| SOC 2 Type I | Cloud control plane | QUEUED Roadmap, first 6 months |
| SOC 2 Type II | Cloud control plane | QUEUED Roadmap, 6 to 12 months |
| HACCP · FSMA · ISO 22000 · FSSC · Codex | Grounding and record-keeping | RUNNING Design basis for retrieval and the audit log |
| Per-tenant isolation | Data, memory, models | RUNNING Design rule |
| On-prem deployment | Recipe-sensitive producers | RUNNING Design option |
Milkeon supports your food-safety records. It is not a certification and it does not replace your HACCP plan.
Enterprise
Standardize autonomy across the group
One policy model, one audit trail and one set of yield, loss, energy and water benchmarks across every plant in the group, with each site's recipes and tolerance models kept private to that site.
- Multi-site plant-edge fleet with shared model governance
- Custom product, composition and tolerance models per site
- SSO and role-based access, named approvers (roadmap, 6 to 12 months)
- Model versioning, validation reports and rollback
- On-prem or private-cloud deployment
- SLAs and outcome-based components on the highest-value workflows
FAQ
Straight answers
The questions plant directors, technologists 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 explicit limits. Every design partner starts in shadow mode, where Milkeon predicts and writes nothing. Assist mode follows, where a technologist approves each recommendation. Bounded autonomy on separation, drying and inspection loops comes after the food-safety and ROI gates agreed up front.
Critical control points are not optimization variables. The pasteurization CCP, the flow-diversion interlock and release decisions stay with your HACCP plan and your people. Milkeon verifies readings, writes them to an immutable audit log and attaches them to the batch genealogy. It supports HACCP, FSMA, ISO 22000 and FSSC records; it is not a certification and does not replace your plan.
Milkeon sits above your existing automation as a supervisory layer. If it stops, the plant keeps running on its current PLC and SCADA control and the interlocks it already has. A Milkeon outage degrades the plant to the way it runs today, not to a stop.
The first phase is shadow mode on one line: Milkeon reads your analyzers, PLC history, lab results and MES while you run the plant as usual, and we measure its predictions against your lab. Recommendations start when accuracy is proven against that baseline, and improve as plant-specific history and technologist corrections accumulate.
Recipes, product specs, tolerance models and supplier data are tenant-scoped and never used to train another customer's models. Retrieval is multi-tenant isolated and permission-aware. On-prem deployment is available for processors protecting sensitive recipes.
You are, the same as with any control strategy. That is why every write is policy-checked, bounded, logged and reversible, every output carries a citation to the recipe or procedure it relied on, and anything above your threshold waits for a named approver. The audit log records the request, the reasoning, the limits applied and the human decision.
As a design partner, on one line, with one hard metric: fat or protein give-away, energy per kilogram of powder, moisture or composition variance, or contamination escapes. We instrument the line in shadow mode, measure the baseline, move to assist mode when accuracy is proven, and expand only when the number holds.
Full multi-module autonomy across the whole plant, long-tail integrations beyond the design partners' stacks, and self-serve onboarding. The first release is one wedge workflow, its connectors, the agent loop with human checkpoints, the review console and the food-safety audit log.
Get started
See these features on your line
Shadow mode shows the perception, prediction and evidence features before any control feature is enabled.