Company
Every short-shelf-life food operation is a central workshop feeding shops against the clock.
Bread is the fastest clock in short-shelf-life food — hours of shelf life, made before dawn, sold or lost by closing. An engine that closes the loop on bread already solves the hardest case in any short-shelf-life operation.
The asset
Both, in layers. The screen a baker uses is the surface; underneath sit four layers — signals, models, decisions, governance — and none of them knows it is working with bread. The trade enters as a declared pack and the same core decides. Baking is the first pack in production; food manufacturing is the second.
What is being built is not an app: it is four layers and one core.
The product a baker sees is the surface. Underneath there is a four-layer stack — signals, models, decisions, governance — and what makes it an asset is that none of the four knows it is working with bread. The trade enters as a declared pack; what decides is always the same core.
What comes in: line-by-line sales, weather at each shop's own coordinates, local calendar and events, surroundings, plant sensors and the order book. Each group is declared once and resolved per trade and per country.
What is learned: one model trained per company, product and location, over sixty-plus declared signals. Its output is not a figure, it is a calibrated band: the interval width is the conformal quantile of that model's out-of-sample error, measured on a held-out window at training and applied at serving.
What is decided: on that band, the core picks the point that minimises each product's expected cost of error — one unit too many against one too few — applies the real oven or line ceiling and writes the trace with the alternatives it dropped.
What is allowed: one evaluator decides whether an agent executes or only proposes, with a fixed precedence. Law outranks the company's policy, and the circuit breaker outranks both: on drift or stale data the engine drops to proposing and no checkbox turns it off.
A vertical declares where its demand comes from, the unit it is measured in, the signal groups it uses and the capacities that constrain it. Baking, food manufacturing and speciality coffee are live. Two more exist on paper only — fresh fish, with its auction as the supply ceiling and the kilo as the unit, and a delivery-only kitchen. All five train a real model and run the same decision kernel in the conformance tests: that is the experiment that proves the four layers do not have bread baked into them.
None of these lines is an intention: each names code that runs today in this platform's tests. What is not true yet is further down, in its own table.
The market map
We sell to one segment today: bakery chains of five to twenty shops with a central obrador, at a published €399 a month and with no prior call. The second layer — industrial, par-baked and pastry — shares the physics and is already written into the code. The third is a topology, not a sector list.
Three layers. You always come in through the first.
This is not a list of sectors we aspire to: it is the order in which they are taken, and each layer opens only with what is already written in the system.
The buyer has a central bakehouse, between five and twenty shops and no coordination layer between them. It is the only segment we enter today, and it is entered with no prior call: published price, public demo, same-day sign-up.
An industrial line and a twenty-shop chain share the physics: dough that ferments, product part-baked here and finished elsewhere, and a shelf life that starts running before it leaves the bakehouse. The size of the equipment changes, the data model does not. The jump is already written into five parts of the system:
The outer layer is described by its shape, not by its sectors: central production, a shop network and same-day shelf life. That is exactly the topology the engine solves, and the proof that the engine does not have bread baked into it is already written:
We do not publish a list of sectors we have not served. The third layer is named by its topology, or by the pack that tests it.
Portability
Because bread is the fastest clock in short-shelf-life food: hours of shelf life, made before dawn, sold or lost by closing. An engine that closes that loop inside one working day — forecast, bake, delivery, invoice — has solved the hardest case. The first vertical is the credential, not the ceiling.
We started with the hardest perishable.
An engine that closes the loop on bread — forecast, bake, delivery and invoice inside one working day — has already solved the most demanding case in any short-shelf-life operation. The country packs, the per-shop models and the dispatch engine are built to outlive the category. The first vertical is the credential, not the ceiling.
Why now
Three clocks started together. Agent orchestration became infrastructure a company can audit line by line. Regulation acquired dates in every market — EU waste targets for 2030, e-invoicing calendars country by country, Spain's Ley 1/2025 obligations in force since 2 April 2026. And capital plus labour push the same way: the giants invest at a rate a regional chain cannot match.
Three clocks that started at the same time.
Agent orchestration is now reliable in production: thirteen coordinated agents, each one's posture — propose or execute — resolved by a single evaluator and by the law that applies to it, and every recommendation traceable to its source data and reversible afterwards. Three years ago this was a conference video; today it is infrastructure a bakehouse can audit line by line.
What was good practice is now a mandatory dated record, country by country. Three fronts, all three with a deadline, and all three touch exactly what this engine decides:
The sector's giants invest at a rate a regional chain cannot match, while the night shift gets harder to cover and energy reprices every quarter. Whoever cannot buy that capacity needs the software to do the coordinating.
The data flywheel
Not pooled data — no customer's rows ever leave its tenant. What compounds is the vertical's engineering: which of the sixty-plus declared signals matter in baking, which fermentation priors need calibrating, what running short really costs on a fast product. Every chain that joins shortens the next one's path to a useful forecast; a horizontal product cannot follow that curve.
Every shop that joins shortens the path for the next one.
Every connected shop brings its own series: line-by-line sales, the weather and the events on its street, the waste at closing, the real fermentation and baking times. That series trains one model per product per shop. What accumulates across customers is not their data — it is the engineering: which signals matter in baking (over sixty), which fermentation priors need calibrating, how much running short on a fast-moving product really costs. Every chain that joins shortens the next one's path to its first useful forecast, without a single data point leaving its tenant. That compounding runs against a horizontal product: the narrower the niche, the faster it learns.
No customer's data is shared with another. What accumulates is the vertical's engineering, never sales or recipes.
Honest roadmap
Live: Spain's fiscal layer with the AEAT, one model per product per shop, the decision kernel, the autonomy evaluator. In progress: TicketBAI homologation, the autonomy console, restore drills, ISO/IEC 27001. Architected or on paper: France and Mexico, two non-bakery packs. Not started: SOC 2. This table is the only place those statuses are stated, and it changes with the code.
What is not true yet, said here and not in a footnote.
A claim we cannot document is not published. These are the rows due diligence asks about most, and the table changes when the code changes, not when the pitch does.
| Item | Status | What is true today |
|---|---|---|
| Forecast model family | Live | Champion per product: Prophet plus an XGBoost residual model, capped at 15–40 features by sales velocity. A global gradient-boosted quantile model runs as a challenger — measured, never served. |
| Interval calibration | Live | The band's width is a conformal quantile of out-of-sample error, measured on a held-out window at training and applied unchanged at serving. Nothing re-fits it between retrains; calibration against live serving error is not built. |
| Vertical neutrality, layer by layer | In progress | Three of the four layers are neutral in the code: the signal registry is declared once and resolved per trade and per country, the decision kernel carries no branch by trade, and governance runs on one evaluator and one proposals table. The model layer works for every declared pack, but its classes are still named after baking; that rename sits on a shrink-only ledger checked in continuous integration. |
| Industrial vertical | Live | The pack is declared and active — order-book demand, kilo grain, line constraint, process signals — alongside master→variant authority, part-baked physics and the cold chain. What does not exist is a plant in production with us: it counts as architecture, never as traction. |
| Portugal, France, Mexico | In progress | Portugal is a real pack for tax, fiscal identity, holidays and geography; its invoicing layer is not built, so issuing in Portugal fails explicitly instead of improvising. France and Mexico are validated on paper against NF525 and CFDI. No customer operating outside Spain. |
| TicketBAI / LROE | In progress | Built and XAdES-signing; foral homologation is in progress. We do not say «certified» until it is. Reading the files your company already issues is available. |
| Plant sensors | In progress | MQTT, OPC-UA and Modbus ingestion plus connectors built for Wachtel and Rational are wired and exercised on the demo stream. No customer plant streams to us today. |
| Hosting and backups | Live | All data at rest in the EU (Germany and Finland) on one provider; backups with the same provider as production, a second-provider copy is a pre-contract item. No per-tenant region pinning. |
| Autonomy per family | In progress | The single evaluator, the auto-mode limits and the precedence — law, breaker, policy, declaration — are written and covered by tests. What is being wired is the screen that edits those policies per family and the feeds that arm the breaker. Until they land, the engine proposes: it never executes on a policy that cannot yet be declared. |
| ERP connectors | In progress | Today we read Microsoft Dynamics 365 Business Central through its own API, plus TicketBAI / LROE files, CSV and the generic ERP export with a declared column mapping. For SAP or Oracle we have no native connector: they are read through that generic export, and we say so before signing. |
| ISO/IEC 27001:2022 | In progress | We are implementing an information security management system. Its scope, policy, risk assessment and a Statement of Applicability covering all 93 Annex A controls are drafted and await management approval. We do not hold the certification and have not engaged a certification body. |
| SOC 2 | Not shipped | Not started. No date committed. |
Market size
We publish a method before a number. Bottom-up: bakery chains of five to twenty shops across Spain and the EU, times a €399 monthly floor, with group and industrial contracts as the ceiling. When the figure is published it will carry the census it was counted from — never a percentage of an analyst report.
A multiplication we can defend.
Bakery chains of 5 to 20 shops across Spain and the EU × €399/month list price as the floor, with group and industrial contracts as the ceiling.
Who builds it
One founder, more than a decade of putting connected systems into other people's buildings: telemetry on a production line, connectivity in tens of millions of devices, a business line built from zero inside a global corporation. BakeWise itself is directed as an AI-agent engineering fleet, behind thirty-plus CI gates and over 35,000 automated tests.
The founder thread.
BakeWise is built by a founder who spent more than a decade putting connected systems into other people's buildings. Telemetry on a production line. Connectivity in tens of millions of devices. A device-to-cloud business line created from zero inside a global corporation. Connected platforms in millions of European households that had to be right every single day, with no technician on site. A bakery is the same problem on a shorter clock, with a margin that does not forgive a bad forecast. That is the thread, and it is why the system looks like infrastructure rather than a dashboard.
Press
With this page and a mailbox. The press kit is not published yet; when it is, it will hold a one-line and a one-paragraph description in three languages, the logo with its clear-space rules, screenshots taken from the demo tenant rather than mocked up, and a data note that never carries a customer's figures.
In preparationPress kit.
Not published yet. When it is, it will contain exactly this:
Press, investors and due diligence: [email protected]