Origin
Where dairy plants lose money and sleep
Talk to a plant director and the picture is consistent. They run largely blind on real-time composition, safety margin, contamination, yield, energy, water and batch genealogy. They over-standardize to avoid off-spec batches, giving away fat and protein on millions of liters. They run dryers conservatively to avoid fouling and moisture misses, burning energy to stay safe. They lean on a workforce that cannot reliably meet rate under HACCP.
The tools they have measure and log. None of them act. That gap, between a plant that is thoroughly instrumented and a plant that is actually controlled, is the opening Milkeon was designed for.
The wedge came from watching where the money and the sleep actually go: give-away on separation and standardization, energy on evaporation and spray-drying, downgrades on fermentation, and recall risk on contamination and seals. Each of those is measurable, tolerance-bound and rate-constrained, which is exactly the shape of a problem autonomy can own.
Workforce
The craft is walking out the door
The best dairy technologists hold the recipe in their heads. Fat and protein get checked by lab sample, not in real time, so plants over-standardize and quietly give away yield every shift. Dryers run conservatively to avoid fouling, burning energy to stay safe. And when a pasteurization or seal problem slips through, it becomes a recall.
That craft is retiring faster than it is being replaced, and the plants that depend on it are being asked to do more, safer, with fewer people.
Our answer is not to replace the technologist. It is to make their corrections the thing the plant learns from: every approval, every override and every note in assist mode becomes a labeled event in a per-plant memory that is versioned, tenant-scoped and never shared across customers.
Simulation
Why a twin has to be as-processed, not as-designed
Plants commit to a run and only discover composition misses, safety-margin problems, fouling or takt bottlenecks after the milk and energy are spent. An as-designed model of the equipment does not help, because the plant never runs as designed: the milk supply changed, the separator fouled a little, the dryer wall is not clean.
Lactwin starts from the live plant state and the coupled physics and biology of separation, heat exchange, fermentation, evaporation and spray-drying, and simulates the batch as it will actually be processed. The result is a recipe and a control envelope the agents can execute, and a record of why the discarded candidates were discarded.
The same machinery generates rare contamination, fouling and drying-upset scenarios that are too rare and costly to collect, always validated against real plant events before a model is allowed to learn from them.
Separation
Give-away is the biggest yield lever nobody measures continuously
Separation and standardization set the economic floor of a dairy plant, and they run on fixed setpoints and periodic lab checks. Between samples, composition drifts. The defensive response is to over-standardize, so fat and protein above spec leave the plant on millions of liters.
Sepion's premise is that give-away is only eliminated if composition is estimated and acted on continuously, not per lab sample. That means fusing inline NIR and mid-IR with the lab calibration and closing the loop on the blend at the minimum safe margin, with every setpoint change and composition outcome in the audit trail.
It is also why the Line tier starts here for most fluid, cheese and powder processors: the baseline is easy to measure and the result is in liters and dollars.
Energy
Drying at the quality-safe frontier
Spray-drying is the plant's largest energy and water consumer, and it is run with margin because the cost of a moisture miss or a fouling event is high. Running conservatively wastes energy and throughput; running hot risks quality and safety.
Aridyn couples a GPU-accelerated CFD and process twin with edge control so the dryer can run closer to the true quality-safe frontier. Spray-drying and evaporation CFD is a classic GPU workload; on CPUs it cannot run fast enough to inform live control or to generate synthetic upset scenarios.
The commercial case follows the physics: energy per kilogram of powder is a number every powder plant already tracks, which makes outcome pricing on verified reduction the strongest value-based case in the portfolio.
Principles
Trust before autonomy
Autonomy in a food plant is earned in stages. Shadow mode predicts and writes nothing. Assist mode recommends and a person decides. Bounded control writes low-risk setpoints inside explicit limits. Critical control points, interlocks and release decisions never move to the agent.
Every output is grounded and cited to the recipe, the HACCP plan or the CIP procedure it relied on. Every action is written to an immutable audit log. Model and prompt changes are gated by golden datasets and model-as-judge evaluations in CI before they reach a plant.
These are the guiding principles in our roadmap, and they are the reason the first release is one wedge workflow with a review console and an audit log, not a plant-wide program.
