← Projects

QBO-F01 · Quobo Labs fleet

Quobo Industrial

FlagshipIn production

Your facility, finally answerable: Quobo's flagship platform models the plant itself — structure, dependencies, live readings, operator knowledge — so a grounded AI can propose work to a named human who always makes the call.

Deployed and operating in production — the story is below.

Chapter one

The problem

An industrial facility — a water plant, a refinery unit, a food-processing line — is run out of three places: a maintenance system that holds work-order paperwork, a SCADA screen that holds live readings, and the head of the most senior person on shift. None of the three talk to each other. The paperwork knows a pump had a work order in March, but not what the pump feeds, why it fails, or what it's doing right now. The historian knows the vibration trend, but not that the same bearing was replaced twice last year. The millwright knows all of it — and retires next spring.

The failure underneath is that the plant's memory has no home. Every maintenance product on the shelf models the work order; the asset is reduced to a reference number on a transaction. Rip the system out and you lose paperwork, not understanding. So the real system of record stays tribal — whiteboards, spreadsheets, a binder of PM schedules — and the standing question no one can answer quickly: if I take this offline, what happens downstream?

The operators who feel this hardest are mid-market industrial — utilities, wastewater, chemical, food processing. Big enough to have SCADA, regulators, and equipment-dense sites. Too small to carry an eighteen-month enterprise implementation. Quobo Industrial was built for them: model the plant, not the paperwork, and give its memory somewhere permanent to live.

The mechanics

How it works

  1. Model the plant

    A facility is built downward, level by level — subsystem, process, asset, endpoint — each carrying its own control narrative, documents, and maps. On top of the hierarchy sits a dependency graph: typed relations across four disciplines, so "this breaker feeds those loads" is a fact in the model, not a memory. A consequence query walks that graph and answers the question every isolation plan starts with: act on this — what's affected? Bulk spreadsheet import brings decades of legacy records in.

    The dependency graph lighting up everything downstream of a breaker before a maintenance window

  2. Teach it what the veterans know

    Classification is bitemporal: what class, make, and model an asset was is queryable as of any date, with an append-only revision history. Failure knowledge — how this class of equipment fails, what to watch for, how the procedure goes — is authored per equipment class by operators, not scraped from a standard. Authored once, it teaches three surfaces: standalone reference, AI narration, and every generated work plan after it.

    A class knowledge packet — failure modes, detection methods, and the authored procedure, side by side

  3. Watch it, honestly

    Edge nodes poll equipment over the industrial protocols plants actually speak, feeding a live telemetry stream. Threshold breaches raise work with severity-mapped priority and resolve themselves when readings recover. And one rule holds everywhere: a stale sensor reads as unknown — not confirmed normal — never as an all-clear. The map, the alarm cards, and the AI all read that same gated truth.

    The operations map — alarm heat where it's real, and honest grey where a sensor has gone quiet

  4. Run the work

    Work orders move through a strict lifecycle with templates, subtasks, and delegation gated by certification — the system confirms the assignee actually holds the required qualification before the handoff completes. Labor hours are cost-stamped as they're logged; parts are cost-stamped as they're drawn. Every closed order deposits a permanent outcome into the asset's history, so repair-or-replace arguments get settled from a ledger instead of a hunch.

  5. Change it like a pull request

    Capital projects get a scratchpad clone of the facility where the team drafts the new structure without touching the live model. Kickoff snapshots the before; merge preview runs a three-way structural diff and surfaces conflicts for a human to resolve; earned-value analysis tracks execution against plan. Git-grade discipline, applied to physical reality.

    A merge preview — before, live, and target side by side, one conflict awaiting a decision

  6. Ask it anything

    OBO, the AI workspace, answers from the plant's own data — the facility model, the open backlog and who's on it, stock levels, live readings — assembled deterministically and scoped to what the asker is allowed to see. Its contract is grounded or honest: cite the real data, or say plainly that the data isn't there. Scheduled queries run standing briefs overnight; AI-synthesized knowledge lands in a review queue where a person endorses or excludes every entry.

    A Monday-morning brief in OBO — open work, assignees, ages, and one honest "unknown" where a sensor went stale

  7. The keystone: it never acts

    The system structurally cannot actuate. It never writes a setpoint, never schedules its own work, never assigns anything to anyone. When a vibration threshold breaches on a pump whose failure knowledge is authored, what arrives is a proposed work order — rationale attached, every step's "why" validated against the authored knowledge, a fabricated citation treated as a hard failure — sitting in an approval queue until a named human approves and delegates it. In a plant where the sign-off chain includes process safety, that refusal is the feature.

Field notes

Would this fit your life?

A maintenance manager who owns the 2 a.m. phone calls

The moment
Overnight, vibration on a pump bearing crossed its authored threshold.
What they do
Finds a proposed work order waiting in the approval queue — failure rationale attached, priority mapped to severity, due date set. Reads the "why," approves, and delegates; the qualification gate confirms the assignee's certification before the handoff lands on their phone.
The payoff
Condition, explained proposal, human approval, qualified hands — and no autonomous action anywhere in the chain.

A technician three weeks into the job

The moment
Assigned a calibration he's never performed on this model of transmitter.
What they do
Opens the AI workspace from the work order's asset and asks for the steps. The draft comes back grounded in the procedure a thirty-year veteran authored for that equipment class — and where the knowledge runs thin, the steps say so plainly instead of inventing certainty.
The payoff
The retiring expert's know-how, authored once, teaching every new hire through every draft after — with an honest flag on everything unverified.

A storeroom coordinator covering two facilities

The moment
A technician draws the last two seal kits against a work order.
What they do
The draw posts to an append-only stock ledger, cost-stamped, and trips the reorder point. She raises a purchase order against the part's supplier; receiving posts the delivery back through the same ledger. Compatibility maps show which replacements fit which assets.
The payoff
Stock truth with a paper trail, a closed reorder loop, and every material dollar landing on the asset that consumed it.

An operations director over a multi-site portfolio

The moment
Monday, 7:30 a.m., before the ops call.
What they do
Asks the AI for a brief on Plant 2 — what's open, who's on it, what's overdue — and gets the answer from the live backlog and facility index, with telemetry claims gated by freshness. His standing weekly digest already ran overnight.
The payoff
The ops call starts from one shared, honestly-hedged picture instead of four people's partial ones.

No varnish

From the lab notebook

Where it stands

In production and operating, run internally by the company that built it — the platform behind quobo.co, deployed with staging, migrations, rollback runbooks, and over a thousand merged pull requests. The core is broad and real: the full facility model with dependency graph and consequence query, the complete work-order lifecycle with certification gates and cost capture, inventory through procurement, version-controlled facility projects, SCADA ingestion with the freshness-gated alarm state, and the grounded AI workspace with its human review queue. This is the reason the lab exists, and it is being flown daily.

What's proven, and what's still gated

The signature chain — a live condition breach becoming an explained, citation-grounded work proposal in a human approval queue — is proven end to end for exactly one equipment domain, on the company's own tenant, and the plan is explicit that proof grows one authored equipment class at a time. AI work-order generation ships behind a flag; the agent runtime is wired but deliberately disabled beyond its reference trigger; the SCADA path has run against simulators and edge hardware but not yet a customer's plant. The docs name each of these plainly — the honesty is part of the engineering culture.

What's next

First customers, onboarded white-glove: the vendor models the facility, imports the legacy records, and authors the first failure knowledge for that plant's equipment classes — which is exactly what turns the AI's value lane on for their domain. Before external doors open, the remaining product-not-internal-tool gaps get closed, external sign-on first among them. The direction doesn't change: the knowledge base deepens per class, the proposals stay proposals, and a human keeps making the call.

Next in the fleet

QBO-F02

Quobo Residential

Pre-launch

The permanent record of a home, from listing to living: one append-only, consent-governed record of everything that ever happened to a house — read and written by owners, agents, contractors, and insurers, where AI drafts and a human always signs.

Continue the tour →