Manufacturing & Supply Chain
Make the plant floor legible before you try to make it intelligent
Most plants already emit the signal needed to predict a bearing failure or explain a scrap spike — it dies in a historian nobody queries. We build the path from PLC tag to planning decision on the IT side of the boundary, leaving the safety layer where it is.
Manufacturing & Supply Chain
Plant-floor signal turned into planning decisions — without ever touching the safety layer
- Now
- 4 applications
- Next
- 2 applications
- Frontier
- 1 application
- Constraints we design around
- 5
Context
Three forces reshaping this sector.
What changed recently enough that the old architecture is now the wrong one.
The second-source build-out is outrunning the instrumentation
China+1 sourcing and PLI-funded capacity coming online across India mean new lines, new contract manufacturers and new tier-2 suppliers every quarter. Each arrives with its own PLC vendor, its own MES — or a spreadsheet — and its own definition of downtime. The cost is not the capex. It is that group-level planning now runs on numbers that are not comparable between plants.
Embodied carbon became a customs calculation
Under the CBAM definitive regime, the EU importer is the party that must declare embedded emissions and surrender certificates — and pushes that evidentiary burden straight back down to the Indian supplier of steel, aluminium, cement or fertiliser, against default values that punish anyone who cannot produce actual figures. Actual figures require energy and material consumption attributed down to a batch. That is a data engineering problem wearing a sustainability badge.
Forty years of fault knowledge is walking out the door
The technician who knows the #3 extruder trips on humid mornings is retiring, and what he knows was never written down. What exists instead is two decades of free-text CMMS work orders, scanned OEM manuals and SOPs in three concurrent revisions. Retrieval over that corpus is finally good enough to be useful — if it is indexed against the asset hierarchy rather than dumped into a vector store.
Applications
What we would build here, in order.
Grouped by how close each one is to production today — not by how impressive it sounds in a board pack.
Build it this quarter
Current models, current data. The engineering is understood and the risk is schedule risk, not research risk.
OEE and downtime attribution that survives an argument
Availability is the number that gets escalated, and it is usually wrong. Micro-stops below the logging threshold disappear, operators pick a reason code from a list of thirty at the end of a shift, and the MES clock has drifted from the PLC clock. The build is a state machine over raw tag transitions, reconciled against work orders, where every downtime event carries the tag evidence that produced it. Arguments end when the evidence is one click away.
Visual inspection running at line rate
Surface defects, missing components, print and seal integrity. The hard part is never the model — it is the fixture, the lighting and the cold start with eleven defective samples. Run anomaly-first, using embedding-distance methods in the PatchCore family, so the line can go live before a labelled defect library exists, then convert to a supervised classifier as the operator review queue fills it. Inference sits on the industrial PC or a Jetson at the line; the cloud receives crops and metrics, not video.
A shop-floor copilot over SOPs and maintenance history
A technician asks why an alarm keeps recurring on line 4 and gets the three prior work orders, the current SOP revision and the OEM manual page — each cited, each clickable. Everything rests on retrieval scoped by asset hierarchy and date window, not on the model. An answer without a citation is refused rather than guessed, and the refusal rate is a tracked metric rather than an embarrassment.
Spare-parts and consumables forecasting
Most SKUs in a plant store move a handful of times a year, which is precisely where ordinary demand forecasting falls over. Intermittent-demand methods — Croston, TSB — combined with criticality weighting beat a generic time-series model. The objective is not forecast accuracy anyway. It is service level per rupee of working capital, with the parts that stop a line treated differently from the parts that do not.
One thing has to change first
Buildable, but gated on something outside the model — labels, a consented holdout, an integration, or a process the business has to agree to change.
Predictive maintenance on rotating assets
Motors, pumps, gearboxes, compressors. Vibration and motor-current signature analysis carry real signal; the obstacle is that a plant with a good reliability record has almost no failures to learn from. Start with unsupervised deviation from a per-asset healthy baseline, route alerts into the CMMS as inspection tasks, and treat the technician’s disposition as the label. The supervised model arrives in year two, trained on labels the system generated itself.
Finite-capacity scheduling and changeover optimisation
MRP assumes infinite capacity, and then a planner repairs it in a spreadsheet. A CP-SAT or MILP model over real changeover matrices, tooling constraints, operator skills and maintenance windows produces a schedule the shop floor will actually run — provided the objective was negotiated before the model was written. Minimising changeover time, maximising on-time delivery and levelling energy load under time-of-day tariffs are three different schedules. Someone has to choose.
Scope it as research, with a gate
Credible and worth a prototype. We would run it behind an evaluation gate and say plainly if it does not clear the bar.
An agentic desk for supply and shop-floor exceptions
A tier-2 supplier misses a shipment, a batch fails QC, a critical spare crosses reorder. Today a planner reconstructs the situation from five systems and emails four people. The agentic version assembles the same picture — open orders, qualified alternates, current schedule, on-hand inventory, supplier risk signals — and proposes a resolution with its reasoning and its uncertainty attached. It proposes. A human commits. Write access to ERP sitting behind an approval step is the whole design, not a feature flag.
Constraints
What makes this sector different.
The part of the build that does not transfer from anywhere else. Every one of these has to be answered in the architecture, not in the project plan.
The OT/IT boundary is not negotiable, and should not be
Plant networks are segmented for good reason — IEC 62443 zones and conduits, Purdue levels, no inbound sessions from the enterprise network down into Level 2. Our read path terminates in the DMZ, at a broker or OPC UA aggregation server the plant owns and publishes outward from. Nothing we build writes to a controller, and the safety instrumented system is outside the scope of every engagement, permanently. An architecture that needs an exception to this is the wrong architecture.
Legacy protocols, and tags that mean nothing
Modbus RTU over a serial gateway, Profibus behind a proprietary bridge, an OPC DA server on a Windows box nobody is permitted to patch. Tags arrive as PLC3_AI_017 with no units, no scaling and no description, because the meaning lives on an HMI screen drawn in 2011. Building the tag-to-asset-to-meaning map is the genuine first deliverable, and it is done with a controls engineer in the room. It is not inferred.
The edge link drops, and the data still has to be right
Plant uplinks fail, edge gateways reboot, clocks drift. Every collector buffers locally and replays — MQTT Sparkplug B birth and death certificates make connection state explicit — and every downstream consumer must be idempotent and tolerant of out-of-order backfill arriving hours late. Systems that assume ordered arrival produce OEE figures that change retroactively, which destroys trust faster than having no figures at all.
Sensor data is abundant; labels are almost nonexistent
Eighteen months of high-rate vibration data and four confirmed failures is the normal case, not bad luck. CMMS work orders record the repair date rather than the onset date, and the free-text reads “found noisy, replaced bearing”. Supervised learning on that is a fantasy. The realistic path is anomaly detection, weak supervision from maintenance text, and a deliberate labelling loop where technicians confirm or reject — designed in from day one, because retrofitting it is far harder.
One global model does not survive eleven plants
The same nominal line runs a different motor vendor in Pune than in Pantnagar, at a different ambient temperature, with different operator practice and a different product mix. A model trained on the flagship plant degrades quietly everywhere else — not with an error, but with precision drifting for a quarter before anyone notices. Per-asset baselines under one shared feature pipeline, a registry that records which model version runs where, and thresholds tunable locally.
Every one of these is discoverable in week one. Most programmes discover them in month three, after the architecture is already load-bearing.
Outcomes
What we design toward.
Targets and typical ranges for this class of system. Each one carries its basis, because a number without one is decoration.
- OEE that becomes visible once downtime is machine-attributed rather than hand-logged
- 5–15 ptsOEE that becomes visible once downtime is machine-attributed rather than hand-loggedTypical range when micro-stops and reason codes move from operator logs to tag-derived events. This is recovered measurement, not recovered capacity — and not a Digiaeon delivery figure.
- Share of unplanned downtime on rotating assets that condition monitoring can address
- 30–50%Share of unplanned downtime on rotating assets that condition monitoring can addressRange reported consistently in public reliability-engineering literature for adequately instrumented assets; the upper end assumes usable failure history exists. Treated as a planning assumption, never a promise.
- Per-part inference budget for edge visual inspection
- < 250 msPer-part inference budget for edge visual inspectionDesign target we hold on a 60-parts-per-minute line, achievable on current edge accelerators. An engineering constraint we design to, stated as such.
- From first tag read to a decision an operator acts on
- 6–10 weeksFrom first tag read to a decision an operator acts onDesign target for one line or one asset class, assuming a controls engineer is available. A multi-plant rollout is a separate and longer programme.
These are design targets and published benchmark ranges, not delivered client results. We do not publish results attributed to clients.
Solutions
What we bring to this sector.
The practices that do most of the work here. Each one is the same discipline applied against the constraints above.
Blueprints
Reference architectures behind this work.
How the systems are actually assembled, with the engineering decisions and their costs stated. Reference designs, not case studies.
Real-Time Decision Platform
A platform that turns database changes into fresh features within seconds, serves a model decision inside a live request, and writes an immutable record of every decision — the inputs, the model version and the policy that applied — so it can be explained, replayed and rolled back months later.
Read the blueprintGrounded Answer Engine
A question-answering system over a company’s own documents that cites every claim to a source span, respects the permissions of the person asking, and refuses when the evidence is not there.
Read the blueprintQuestions
The awkward questions.
The ones a technology committee actually asks, answered the way we would answer them in the room.
Next step
Bring us the constraint you keep working around.
A 45-minute working session. We'll tell you what we'd build, what we'd not build, and roughly what it costs. No deck.