An intelligence company that happens to clean
Commercial cleaning is priced by the hour and managed from memory. Rotas live in a phone, quality is whatever the supervisor last saw, and nobody can say with confidence what actually happened on a floor at two in the morning. That is not a labour problem. It is an information problem, and it is the reason the work has resisted automation for forty years.
We sell efficient clean results, not hours. What we are building is the reasoning layer that decides how they get clean. Machines are already on site and improving fast. What nobody has built is the layer above them — the one that knows the building, holds the plan, and decides shift by shift whether a zone is better served by a person or a machine.
Space
Metric, zoned, semantically labelled floor models of every building we clean — not a floor plan in a drawer, a live model that machines rebuild on every run.
Time
One event chain from bid and quote through rota, shift and sign-off. Every state change in the business is the same kind of object.
Work
A single task graph covering people and machines together, with each zone assigned to whichever is genuinely better at it that night.
Judgement
A reasoning layer that works from live state rather than a fixed rota, and acts through a permissioned, audited gateway.
How the intelligence compounds
Sense
Machines return telemetry, coverage and fresh SLAM floor maps. The cleaner app returns arrivals, exceptions and photo evidence from the people on site.
Instrumented by defaultDecide
The orchestrator reasons over live state — what is due, who actually turned up, which zones suit a machine, what changed since last night.
Shift by shiftAct
Instructions go out through a permissioned gateway. Machine commands are reversible and safety-interlocked, so they run first; anything affecting a person keeps a human in the loop.
Audited, reversibleLearn
Every completed zone tightens the hours model that prices the next contract. The estimate stops being a guess and becomes a measurement.
The flywheelOne layer over every floor
Five layers, one event chain. Surfaces and devices produce events, the stream carries them, state views derive truth from them, a reasoning layer decides what happens next, and a permissioned gateway executes it. It runs robots from any manufacturer.
Digital twin
Surveyed once, then kept current by the machines themselves — every run returns fresh SLAM map data. The twin does two jobs: it prices contracts from cleanable square metres by surface type, and it assigns zones to people or machines geospatially.
Fleet management
Machine commands are reversible and safety-interlocked, so the orchestrator can start, pause and reroute a unit on its own authority. Supervisors see faults before a client does, and every action is logged against the site it touched.
| Unit | Site | Task | Coverage | Battery | Status |
|---|
COVERAGE THROUGH THE SHIFT — SQ.M / HOUR
EXCEPTIONS
The architecture
Action gateway
Permissioned, audited execution. Every action the system takes passes through one place, with machine commands cleared to run autonomously and people-affecting actions held for approval.
Orchestrator
The reasoning layer. It reads live state rather than a fixed schedule and decides what happens next — which zones, which units, which people, in what order, tonight.
State views
Derived truth. What is due, who is actually on site, which zones are clean and which are drifting — computed from the stream rather than typed into a spreadsheet by a supervisor.
Event stream
One chain carrying the whole business: bid, quote, contract, rota, arrival, task, exception, sign-off, invoice, renewal. Every state change is the same kind of object.
Surfaces & devices
The cleaner app, the manager console, robot APIs across any OEM, site sensors, and the building twin. This is where events are born and where instructions land.
Building the first orchestrated cleaning company
Cleaning is one of the last large industries still run almost entirely on memory. The work leaves no trace, so it cannot be measured, so it cannot be improved. We are replacing that with an instrumented operation and a workforce that supervises machines instead of pushing them.
TODAY — RUN ON MEMORY
The work happens, then vanishes
- The night shift produces no data, so last night is unknowable by morning
- Quality is whatever the supervisor saw last — nothing is auditable
- Rotas live in a phone, and the knowledge leaves with whoever was holding it
- Priced by the hour and squeezed at every tender, with nothing to defend the margin
TOMORROW — INSTRUMENTED & ORCHESTRATED
Every shift observed, reasoned over and recorded
- Crews stay and earn more, supervising fleets instead of pushing machines
- Every square metre timestamped, traced and verifiable on demand
- One live picture across every site, and each site added sharpens the model
- Contracts priced from measured square metres rather than an estimator's instinct
Instrument
Put the cleaner app, robot APIs and a live building twin on every site, so the work produces data as a by-product of being done.
Reason
Hand that live state to an orchestrator that decides what happens next, rather than replaying a rota written six months ago.
Compound
Every zone cleaned sharpens the model that prices the next contract and plans the next shift. The advantage widens on its own.
Anyone can clean a floor. Almost nobody knows what happened on it.