Industrial production

The same engine, architected for the industrial line.

A plant does not need another product. It needs the engine to understand its structure: who owns the formula, what state each dough is in, and which signal arrives from which machine.

That engine — the BakeWise Biga Engine — is the same one running today in chains and groups. This page separates two things industrial pitches usually blend: what is already in production, and what is architected and not yet built.

Request an infrastructure briefing

What this page is

This page describes architectural capability that is verifiable in the code, not industrial deployments under way. We publish no plant customers or references.

R&D → scale-up

From the bench to the line: the scale-up is predicted, not guessed.

Developing an industrial product doesn't end at a formula that wins a tasting: it ends when that formula runs on the line without surprises. The engine carries that path —lab, pilot, scale-up, industrialised— as trial stages. At the step to scale-up it predicts the fermentation minutes and the line speed from the plant's own calibration, not from a generic table. The prediction is checked against the first real batches.

Predicted scale-up
When a trial branch moves to a specific line, the engine predicts the fermentation minutes from that site's calibration and the line speed from that equipment's throughput per hour. Predicted-versus-actual fills in as the batches land: the scale-up stops being a leap of faith.
Trial stages and funnel
Every trial lives in a stage —lab, pilot, scale-up, industrialised— and the per-stage funnel says how many sit in each and where they stall. The question “have we tried this already?” is answered by searching by ingredient, not from memory.
Co-creation with the customer
A development brief arrives with its constraints —target claims, excluded allergens, cost and shelf life— and links to the trials that resolve it, the samples sent and the verdict. Whoever raised it follows it from their own tab in the portal.

Plant vocabulary

Does the engine know a line from an oven?

It knows the objects a plant names. Seven words from daily use on a line, each with the object in the engine that answers it. Where that object is only half there, the row says which part is missing. Every status is a claim about code, not a roadmap.

The terms a plant floor uses, and what backs each one

Term	What answers it	Status
OEE (availability × performance × quality)	Computed per equipment and rolled up per site, with a worst-to-best ranking and a "no data" sentinel so a site with no readings can never read as a healthy site.	In production
Line	The line is now an object in the engine. Every machine belongs to a line of its site, in order, and the live synoptic draws it with readings, running batches and state. OEE is still computed per equipment and rolled up per site; line-level OEE is designed on that base, not built.	Partial
Changeovers	The engine records the equipment cleaning cycle and flags any pair of consecutive batches with conflicting allergens when no cleaning is on record. What it does not yet carry is changeover time and loss accounting.	Partial
Giveaway	Every batch carries actual yield against plan and waste with its reason, so batch-level deviation is measured. A per-unit measure against a target weight is not built.	Partial
Yield	Actual against planned yield per batch, with per-batch consumption of each ingredient and its variance, and actual against expected cost on the same record.	In production
Cold chain	Per-zone thresholds — freezing −25 to −18 °C, refrigeration 2 to 4 °C — logged per equipment, auto-alerted out of range, with the full series available as audit evidence.	In production
Recall	Two phases: synchronous quarantine, then background notifications, with a distribution trace that lists every destination and separates exact from probable certainty. No destination is reported as notified.	In production

ISA-88

What does a plant declare that a bakery does not?

Authority over a formula, stage by stage. A bakery has one recipe; a plant declares who owns the master and which stages a line may change at all. The engine carries that as a first-class object, and the server refuses an edit outside the declared limits rather than accepting it and correcting it later.

Master→variant authority: who owns the formula

The master recipe is the general recipe, owned by the site that defines it. Each line or site may hold its control recipe as a variant. It shadows the master in the effective read, never mutates it, and a master cannot be deleted while variants exist. Authority is declared stage by stage and the server evaluates it fail-closed — an out-of-limit edit answers 422 rather than being accepted and corrected later.

Fixed
The stage is untouchable. Fermentation, critical temperatures and times stay as the master's owner defined them.
Adjustable within limits
The line tunes inside a range declared on the stage itself. Outside the range, the server refuses.
Optional
The stage may be skipped in the variant without breaking alignment with the master's sequence.
Free point
The variant may insert its own stages at that point in the process, and only at that point.

State physics

Par-baked is a state, not a label

Prepared dough, frozen product, par-baked and fully baked are first-class inventory states. Every transformation declares how expiry is recomputed — from the transformation date or preserving the original — and shelf life travels with the lot, not with the product record.

Fermentation-kinetics engine

Proof-time prediction by Arrhenius/Q10, invertible towards dough temperature or yeast dose, calibrated per site against that site's own logs, with ± bands derived from the residuals. When a site is uncalibrated, the answer says so.

Signal ingestion

From the machine to the engine

Protocols
MQTT, OPC-UA, Modbus and REST, with scheduled equipment polling and a batch ingestion endpoint for plant volumes.
Oven adapters
Beyond the generic protocols, two adapters are written against specific Rational and Wachtel equipment.
Anomaly detection
Z-Score, IQR and Isolation Forest over the equipment series, with the method and the window visible in the response.
Predictive maintenance
A Weibull fit over the equipment's failure history to estimate failure probability per window, not a threshold alarm.
One model per product per site
Each demand model belongs to exactly one (company, product, site) triple. There is no global model shared across plants.

Product development

The laboratory, inside the same engine

A trial compares arms of a formula with the costing, the nutrition and the allergens of each. That comparison also flags which comparative claims —reduced, no added sugars— each arm meets against the control arm. On top sit the three things that turn an opinion into a file: bench measurements, tasting and shelf life.

Laboratory results
Water activity, moisture, specific volume, crumb firmness: a measurement is a value with its unit, its method, its technician and its moment, tied to the arm that produced it. Entered by hand or imported by CSV, with the source recorded — the first question anybody asks of a number in an audit.
Sensory panels
At least two tasters per panel, an aggregate derived on close and immutable afterwards. The aggregate travels into the arm comparison, beside cost and nutrition: what the room thinks sits beside what the instrument says.
Shelf-life studies
The design declares the time points — D+0, D+3, D+7, D+14, D+28 — and the chamber conditions. It closes out arm by arm with the day the product went off and its failure mode. An arm that survives the whole design closes with no failure day, which is a result and not a gap.
Industrial palette
180 processing aids, enzymes, emulsifiers, hydrocolloids and laminating fats with their E-number, their functional class and the allergens certain from the identity of the material. EU-authorised additives only: a palette a technologist can dose is a palette that is legal in the jurisdiction the plant sits in.

Configuration, not forks

The industrial shape is a vertical of the engine

The engine carries a vertical marker per company. In the industrial shape the retail surfaces — POS, online store, shop pages — are absent and the vocabulary shifts from obrador and shop to plant and site. The platform sets it, and a company can never switch its own vertical. If resolution fails the value falls to the retail side, because the vertical describes product shape and never decides a permission.

Same tables. Same services. Configuration, not forks.

The honest split

What is built, and what is only designed?

Two columns, and the distinction is not rhetorical. The left one opens in a technical session over real data; the right one is design, not software. The right column shrinks as capabilities land and is never quietly deleted, so what it names today is what is genuinely missing today.

What is in production and what is architected

In production today

Sensor ingestion over MQTT, OPC-UA, Modbus and REST, with scheduled polling
OEE per equipment and rolled up per site, with ranking and a "no data" sentinel
ISA-88 stage-by-stage authority between master recipe and variant, evaluated fail-closed
Prepared-dough, frozen, par-baked and baked states, with per-state expiry and declared transformations
Fermentation kinetics with per-site calibration and ± bands derived from the residuals
One demand model trained per product per site
Two-phase recall with a destination-by-destination distribution trace
Cold chain with per-zone thresholds, out-of-range alerting and the full series for audit
R&D trials with the formulation frozen into a version, and cost, nutrition and allergen comparison across arms
Sensory panels with a two-taster minimum, an aggregate derived on close and immutable afterwards
Costing at formulation level, with labour and energy at the real tariff
ERP export ingestion with a declared column mapping, line reference and process order included
Laboratory results per arm, by hand or by CSV import, with method, technician, moment and source recorded
Shelf-life studies with declared time points and arm-by-arm close-out, carrying the failure day and mode
An industrial palette of 180 materials with E-number, functional class and identity-certain allergens, EU-authorised only
SAML 2.0 single sign-on, SCIM 2.0 user provisioning and department-scoped permissions
OTIF over the trailing window for B2B dispatch, filtered by the viewer’s own site scope
A published sensor integration contract: a reading is one machine at one instant, and an integration that splits it gets an explicit rejection
Live line synoptic: every machine on its line, in order, with readings, the running batch and its state, on the plant wall
Shelf-life prompt: when a study's declared time point comes due with no measure, a prompt lands in the R&D queue
Predicted trial scale-up: from lab to pilot to line, with the fermentation minutes and line speed derived from the plant's own calibration
Trial stages —lab, pilot, scale-up, industrialised— with a per-stage funnel and search by ingredient
Comparative nutrition claims —reduced, no added sugars— computed against the trial's control branch
Customer development briefs: from the brief to the linked trials, the samples and the verdict, with a tab in the customer portal

Architected, not built

Per-line OEE boards: the line is now a first-class object with its live synoptic, but OEE is still computed per machine and rolled up per site
Changeover time and loss accounting
Per-unit giveaway against a target weight
The customer dimension in forecasting: training is per product per site today, not per account or channel
Distribution networks deeper than one echelon: the topology is a two-level forest today
A native SAP adapter: the available path is export files with a declared column mapping
Novel-food regulatory screening and labelling divergence between markets

Evaluating the engine for a plant?

A technical session on the material above: stage-by-stage authority, signal ingestion, state physics and the limits in the right column. With the people who write the engine.

Request an infrastructure briefing See the industrial demo

The briefing is not a sales demo and carries no purchase commitment; the demo opens on its own, with the data of an example plant.