Origin
Where bakeries lose margin they cannot see
Talk to a plant director and the same costs surface. Depositors overshoot the target weight to avoid ever going underweight, so giveaway bleeds out gram by gram, millions of times a day. Ovens run fixed profiles while incoming dough varies, so under- and over-bake downgrade product and shorten shelf life. Baking is the biggest energy line in the plant.
Then there is the fear: one foreign body, one allergen slip or one seal failure becomes a recall that costs millions and breaks customer trust overnight.
The tools a bakery has measure these problems. None of them fix the root cause, because inspection is read-only and equipment holds fixed profiles. That gap, between a line that is thoroughly instrumented and a line that is actually controlled, is the opening Doughex was designed for.
Workforce
Dough is alive, and the people who can read it are retiring
A head baker can look at dough and know it is proofing too fast. A process engineer can hear an oven and know the bake is drifting. That craft lives in a few people's hands and instincts, and it is aging out of the industry faster than it can be replaced.
Meanwhile the plant runs around the clock, every shift, on lines that never stop, and the craft cannot be cloned across every line and every site.
Our answer is not to replace the baker. It is to make their corrections the thing the plant learns from: every approval, override and note in recommend mode becomes a labelled event in a per-plant memory that is versioned, tenant-scoped and never shared across customers.
Simulation
Why a twin has to be as-baked, not as-designed
Bakeries commit to a run and only discover colour misses, weight drift, collapsed proof or takt bottlenecks after the dough and the energy are spent. An as-designed model of the oven does not help, because the line never runs as designed: the flour lot changed, the proofer humidity drifted, the oven band is a little slower on the left.
Bakesim starts from the live plant state and the coupled physics of oven heat and airflow, proof heat and humidity, and dough rheology and deformation, and simulates the product as it will actually be baked. The result is a profile and a control envelope the agents can execute, and a record of why the discarded candidates were discarded.
The same machinery generates rare foreign-body, under- and over-bake and deformation scenarios that are too rare and costly to collect, always validated against real plant events before a model is allowed to learn from them.
Depositing
Giveaway is the biggest yield lever nobody controls continuously
Depositing and dividing portion every unit the bakery makes, and they drift with dough rheology, temperature, hydration and machine wear. The defensive response is to overshoot the target so nothing ever goes underweight, and that margin leaves the plant on every unit.
Gramex's premise is that giveaway is only eliminated if portion weight is predicted and acted on continuously, not corrected by hand between checkweigh readings. That means fusing the checkweigher with dough state and depositor feedback and trimming the setpoint at the minimum safe margin, with an underweight guardrail and every trim in the audit trail.
It is also why the Line tier starts here for many biscuit, cookie and cake bakeries: the baseline is easy to measure and the result is in grams and dollars.
Energy
Baking at the quality-safe frontier
The oven is the plant's largest energy consumer, and it is run with margin because the cost of a pale loaf or a wet crumb is high. Running conservatively wastes energy; running hot risks over-bake, short shelf life and scrap.
Ovenix couples colour and moisture sensing with a GPU-accelerated oven twin so the profile can run closer to the true quality-safe frontier. Oven CFD, heat and airflow across zones, 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 unit is a number every industrial bakery already tracks, which makes outcome pricing on verified reduction one of the strongest value-based cases in the plant.
Principles
Trust before autonomy
Autonomy in a food plant is earned in levels. Observe mode predicts and writes nothing. Recommend mode proposes and a person decides. Approve-to-apply prepares the write and waits. Bounded control writes low-risk setpoints inside explicit limits. Kill steps, metal detection, allergen changeover and release decisions never move to the agent.
Every output is grounded and cited to the recipe, the proof or bake spec or the HACCP plan 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 programme.
