AI for manufacturing
AI for manufacturers
Downtime predicted, defects caught, and decades of operator knowledge captured before it retires. 5x ROI in 30 days, or we work free.
- Job shops
- Metal fabricators
- Food processors
- Injection moulders
- Contract manufacturers
Teams we build for
- Hoyes Michalos
- Nurse Next Door
- Fedi
- UBC Sauder
- Merchant House Capital
- Picton Investments
- Campbell Froh May & Rice LLP
- Barnakl
- Hungerford
- Breez
Teams we build for
- Hoyes Michalos
- Nurse Next Door
- Fedi
- UBC Sauder
- Merchant House Capital
- Picton Investments
- Campbell Froh May & Rice LLP
- Barnakl
- Hungerford
- Breez
Teams we build for
- Hoyes Michalos
- Nurse Next Door
- Fedi
- UBC Sauder
- Merchant House Capital
- Picton Investments
- Campbell Froh May & Rice LLP
- Barnakl
- Hungerford
- Breez
Proof, honestly
No published manufacturing case study yet, and we say so. Founding clients get priority scheduling, direct founder involvement, and a co-published case study once the numbers are real. The 5x ROI guarantee applies from day one.
We guarantee 5x ROI inside 30 days of deployment, in writing, measured against a baseline you sign before we build. If the system misses the bar, we keep working for free until it clears.
Unplanned downtime is the most expensive hour in the plant.
The bearing that started singing on Tuesday becomes the breakdown that kills Thursday's schedule. The warning was in the vibration data the whole time. Nobody was looking, because watching trends is nobody's job at a plant this size.
Quality escapes cost the relationship, not just the part.
Manual inspection at line speed misses what an eye cannot hold for eight hours. The escape gets found on the customer's dock, and the corrective action request arrives with next quarter's volumes riding on the answer.
Your best troubleshooter has a retirement date.
Decades of what-that-noise-means lives in one head. The binder by the machine stopped being updated years ago, and when he is on vacation the line waits for Monday. Nothing in the plant captures what he knows, and everyone knows it.
The paperwork ships later than the parts.
The certs, the first article report, the COA. The customer's dock wants the documents before the parts count as delivered, assembling the package means an afternoon in job folders, and the invoice waits with it.
The ERP says one thing. The floor says another.
Job status lives on the traveler and the whiteboard, and the ERP catches up at day end if the data entry happens at all. Quoting, scheduling and costing all inherit the gap, and the promise date is a guess wearing a report.
Every NCR is an essay nobody has time to write.
The nonconformance gets dispositioned in the morning huddle and written up never. The 8D goes out late, the same defect comes back in the spring, and the audit finding is sitting right there in the paperwork you did not have time to do.
Manufacturing, before and after
The manual path is dashed: Breakdowns arrive without warning, Inspection misses at line speed, Know-how walks out the door. The system path replaces it, and a person approves before anything ships: Failure patterns flagged early, Defects caught automatically, Tribal knowledge searchable.
Before: by hand
- 01Breakdowns arrive without warninghuman
- 02Inspection misses at line speedhuman
- 03Know-how walks out the doorhuman
After: the system
- 01Failure patterns flagged early
- 02Defects caught automatically
- 03Your approvalhuman
- 04Tribal knowledge searchable
The research
US manufacturing could need 3.8 million new workers by 2033 and leave 1.9 million roles unfilled. 65% of manufacturers call talent their primary challenge (Manufacturing Institute and Deloitte, 2024, US projection).
The Manufacturing Institute and Deloitte · 2024 · Online survey of more than 200 US manufacturers plus executive interviews and labour supply and demand modelling, published April 2024
The numbers in manufacturers
US manufacturing could need 3.8 million new workers by 2033 and leave 1.9 million roles unfilled. 65% of manufacturers call talent their primary challenge (Manufacturing Institute and Deloitte, 2024, US projection).
The Manufacturing Institute and Deloitte · 2024 · Online survey of more than 200 US manufacturers plus executive interviews and labour supply and demand modelling, published April 2024
GE reports a 20% reduction in unplanned downtime from AI predictive maintenance (GE via UITG, 2025).
GE (via UITG industry roundup) · 2025
Computer-vision inspection reaches 98-99% defect-detection accuracy as an industry benchmark (Voxel51, 2025).
Voxel51 · 2025
What we build for manufacturers
Keep the line running
Predictive maintenance.
Sensor data watched for failure patterns, work orders raised before the breakdown, backtested against your own downtime history first.
Tribal-knowledge copilot.
Troubleshooting captured from your veteran operators, searchable from the floor in a technician's own words.
Prove the quality
Computer-vision quality inspection.
Defect detection at line speed on the repeatable checks. People keep the judgment calls, and the line keeps its pace.
Cert package and FAI assembly.
Mill certs, inspection records and the COA assembled against job and lot, gaps listed instead of discovered on the customer's dock.
NCR and CAPA first drafts.
The nonconformance written up from inspection data with disposition options laid out. The quality manager decides, the record files itself against the lot.
Supplier document and cert chase.
Material certs and COAs requested, tracked and filed against every incoming lot, with hold recommendations where paperwork is missing.
Feed the office
Production scheduling and changeover optimization.
Schedules built from open orders, real constraints and floor state instead of the whiteboard.
Maintenance and downtime reporting.
Causes, costs and trends assembled automatically, so the next capex case arrives with its evidence attached.
Where the data comes from
MES and ERP
The job records, BOMs, routings, POs and lot data, whether that is Genius ERP or ProShop in a job shop, SyteLine on a fab floor, or BatchMaster in a food plant. This is where a part number becomes a job and where the cert package finds its spine, and the traveler layer beside it holds what actually happened on the floor, on paper more often than anyone admits.
Where it stops. Read only, and that is the whole design. No job closes, no inventory transaction posts, no cost adjusts. Drafts and proposals flow out of the system, a person posts what they accept, and the ERP keeps exactly the write access map it has today.
Sensor history
The historian and SCADA layer, Ignition or FactoryTalk in most plants, plus the vibration and temperature feeds and the CSV exports the CNC controls have been quietly producing for years. This is where the failure that surprises everyone on Thursday was already visible on Tuesday, and where prediction is backtested against the plant's own past instead of a vendor's brochure.
Where it stops. The data lane only, never the control lane. Reads come one way from the OT side, nothing reaches back toward a PLC or a setpoint, and no output of this system can actuate anything. A work order proposal is as far downstream as it goes.
Maintenance logs
The CMMS where one exists, Fiix or MaintainX in shops this size, and the paper work orders and the binder by the machine where one does not. Two decades of what-that-noise-means is scattered through closed work orders and one veteran's memory, and this is the corpus the troubleshooting copilot is built from.
Where it stops. Work orders arrive in the queue as drafts with evidence attached, and the planner owns priority and schedule. Nothing in the log is used to score or grade a technician, which is a design promise, not a policy memo.
A quality manager's Thursday at a job shop
The customer wants the paperwork before the parts.
humanAt 10:20 the email lands. Their receiving dock has the order on hold pending the material certs and the first article report. She asks for the lot's document package instead of blocking off her afternoon for the job folders.
The document package assembles from the job record.
Mill certs, the certificate of conformance, in-process inspection records and the first article report are pulled together against the job number and lot. Every document carries where it came from. One item comes back flagged, not found, instead of quietly absent.
She catches the heat number mismatch.
humanOne mill cert's heat number does not match the traveler. She pulls the right cert from receiving records, swaps it in and approves the package for send. Nothing left for the customer until her approval, and the flag is why she looked.
Tuesday's NCR arrives drafted, not blank.
The undersized bores from Tuesday are written up from the inspection data, the affected quantity counted, the containment checklist opened and the disposition options laid out. The write-up that never happens in a busy week is already done. It dispositions nothing.
Disposition stays a two-person call.
humanShe and the production manager decide, rework against the print. The system files the record with both names on it. Use-as-is, rework or scrap was never its call to make.
The missing mill certs get chased at the source.
Two raw material lots arrived this week without certs. Requests go to both suppliers with the PO lines referenced, and the receiving clerk gets a hold recommendation on the uncertified material. The hold itself stays a human click in the ERP.
The press brake asks for maintenance before it fails.
The vibration trend on the press brake matches a pattern from the sensor history, and a work order draft lands in the maintenance queue with the evidence attached. The planner will decide when. The machine keeps running while people schedule the intervention instead of surviving it.
The day's records land where the auditor will look.
The cert package, the NCR and its disposition, and the inspection records all index against job, lot and machine. The next customer audit request is a search. The next mock recall is a search. Nobody rebuilds today from memory in six months.
Who this is built for
The plant tells you before the customer does
Today the bad news arrives from outside, a rejected shipment, a corrective action request, a breakdown that eats the week. The system watches the sensor history, the quality stream and the job records and raises its hand first. You still decide what to do about it. You just stop being the last to know.
Every lot ships with its paperwork attached
The certs, the inspection records and the COA assemble from the job record while the parts are still on the truck dock, and what is missing is listed as missing instead of discovered later. The customer's receiving clerk stops being the person who finds your gaps. Audit prep shrinks to reading.
The symptom finds the fix that worked last time
The knock in the hydraulic unit was diagnosed before, by the millwright who retires next year, and the fix lives in his head and a work order comment from years back. Described in a technician's own words, the symptom pulls up that record with the machine, the parts used and what actually worked. His last two years become the plant's record instead of his leaving gift.
The whiteboard and the ERP finally agree
The schedule you publish is only as good as yesterday's data entry, so you walk the floor to find the truth before you promise a date. When traveler status, machine state and open orders read into one picture, the promise date comes from the plant as it is this morning. You still own the sequence and the overtime call.
What stays human
- The customer wants the paperwork before the parts
- She catches the heat number mismatch
- Disposition stays a two-person call
The steps the day below leaves to a person, by design.
The economics
Before and after economics
Line
- Documentation and reporting hours in scope
Before
At the page defaults, 8 people spending 8 hours a week on certs, travelers, NCR write-ups and downtime reports is 64 hours, $3,520 a week at $55 loaded cost
After
The calculator below reprices it with your own headcount and rates, and the 5x ROI guarantee is measured against a baseline you sign before we build
- The cert package on a customer hold
Before
An afternoon in job folders while the shipment sits on their dock and the invoice waits with it
After
Assembled from the job record for a quality manager's review, qualitative until a founding client signs the numbers
- Unplanned downtime
Before
The breakdown announces itself, and the schedule absorbs it
After
Failure patterns backtested against your own downtime history before any schedule changes, with precision and recall reported as separate numbers
- The retiring millwright's knowledge
Before
In one head and a binder that stopped being updated years ago
After
Indexed and searchable from the floor, and the value shows the first week he is away
No manufacturing client outcome numbers exist yet, and this table invents none. The only figures are this page's calculator defaults, 8 people at 8 documentation and reporting hours a week at a $55 loaded hourly cost, which is 64 hours or $3,520 a week in scope. Every other row is a before and after state, not a measurement. The founding tier exists because we would rather publish your numbers than borrow someone else's.
Run your numbers.
Your operations
Savings use the low end of our hours-reclaimed range. The math is conservative on purpose.
The math
Calculated at the low end of every range.
Every first build is covered in writing: 5x ROI in 30 days. Or we work for free.
The objections
Our ERP is fifteen years old and half the plant runs on paper travelers.
Nobody is migrating you off E2. The build reads job records where they are and scans the travelers the floor already fills in, because the traveler is usually the truest record in the building. If a customer's cert request cannot be answered from what exists, you hear that in the audit, before a build, not after one.
The last software rollout died on the floor. Operators will not touch a screen mid-shift.
We do not ask them to. The floor-facing piece takes a spoken question or the photo of a traveler, and everything else runs on records the plant already produces. In the adoption engagement linked on this page, usage of tools the company had already licensed climbed from a small minority to nearly everyone, and it climbed by removing steps, not adding screens. The same discipline applies here or the build fails, and we say so up front.
If your system writes into the ERP, it can corrupt job costing, inventory and everything downstream.
It cannot, because it never writes there. Work orders, NCR drafts, cert packages and hold recommendations all arrive as proposals, and a person posts what they accept. The ERP stays the system of record with exactly the write access it has today, which is the arrangement your accountant and your auditor would design.
Our process parameters and recipes are the business. One leak and we train our own competitor.
Feeds, speeds, cycle recipes and yields stay in your accounts, scoped per line and per plant at the database layer, under no-training API terms. Nothing about your process improves anyone's model or any other client's build. Where controlled goods apply we go further and the data never leaves your environment at all.
A false alarm costs us a changeover. Crying wolf is worse than the breakdown.
Agreed, which is why false alarms are priced into the build instead of hidden in an accuracy number. Failure prediction is backtested against your own downtime history, precision and recall are reported separately, and the maintenance schedule does not change until you have seen what the alert rate would have been last year. If your history cannot support a trustworthy alert, we tell you that instead.
Built around your rules
| Regime | What it demands here | How the system complies |
|---|---|---|
| Customer quality standards (ISO 9001, AS9100, IATF 16949) | Records, traceability and signoff that survive your registrar's surveillance audit and your customer's flow-downs. | The system generates its records inside your existing QMS procedures, who inspected, against what rev, dispositioned by whom. We certify nothing and claim nothing. We build to pass the audits you already face, and the auditor sees better paperwork, not a black box. |
| Safe Food for Canadians Regulations and HACCP | Lot traceability one step forward and one step back, producible on demand when CFIA asks, with the preventive control plan lived rather than laminated. | Lot records, COAs and supplier certs index against every batch as they arrive, so a mock recall becomes a search instead of a weekend. Your preventive control plan stays yours, and every disposition stays with your QA lead. |
| WorkSafeBC and OHS | Cameras and sensors on a floor where people work must watch the process, not the people, and safety calls stay human. | Floor vision points at product and process and is designed around worker privacy. Nothing in the data grades an operator, and no safety decision is delegated to a model. Supervisors keep the calls they have today. |
| Controlled goods and export controls | Where drawings or technical data fall under Canada's Controlled Goods Program or export control lists, the data stays inside the registered environment and in front of authorized people only. | Those builds deploy entirely inside your environment, access is scoped per person at the database layer, and controlled technical data never leaves. Where that rules out a capability, the capability is what goes. |
| PIPEDA | Employee personal information, the names on work orders, training records and shift logs, handled under Canadian privacy law. | The system indexes machines, lots and jobs, not people. Where a name rides on a work order it stays inside your accounts, and no individual operator is profiled or scored by anything we build. |
The regimes that govern manufacturing, what each demands, and how the system complies
Regime
- Customer quality standards (ISO 9001, AS9100, IATF 16949)
What it demands here
Records, traceability and signoff that survive your registrar's surveillance audit and your customer's flow-downs.
How the system complies
The system generates its records inside your existing QMS procedures, who inspected, against what rev, dispositioned by whom. We certify nothing and claim nothing. We build to pass the audits you already face, and the auditor sees better paperwork, not a black box.
- Safe Food for Canadians Regulations and HACCP
What it demands here
Lot traceability one step forward and one step back, producible on demand when CFIA asks, with the preventive control plan lived rather than laminated.
How the system complies
Lot records, COAs and supplier certs index against every batch as they arrive, so a mock recall becomes a search instead of a weekend. Your preventive control plan stays yours, and every disposition stays with your QA lead.
- WorkSafeBC and OHS
What it demands here
Cameras and sensors on a floor where people work must watch the process, not the people, and safety calls stay human.
How the system complies
Floor vision points at product and process and is designed around worker privacy. Nothing in the data grades an operator, and no safety decision is delegated to a model. Supervisors keep the calls they have today.
- Controlled goods and export controls
What it demands here
Where drawings or technical data fall under Canada's Controlled Goods Program or export control lists, the data stays inside the registered environment and in front of authorized people only.
How the system complies
Those builds deploy entirely inside your environment, access is scoped per person at the database layer, and controlled technical data never leaves. Where that rules out a capability, the capability is what goes.
- PIPEDA
What it demands here
Employee personal information, the names on work orders, training records and shift logs, handled under Canadian privacy law.
How the system complies
The system indexes machines, lots and jobs, not people. Where a name rides on a work order it stays inside your accounts, and no individual operator is profiled or scored by anything we build.
Process data, drawings and recipes stay inside plant-controlled systems. Nothing proprietary leaves your accounts or trains a model.
How the system is built for manufacturers
See the full capability mapRetrieval
Maintenance logs, SOPs and the troubleshooting knowledge that currently lives with two people, captured and indexed. A technician describes the symptom in their own words and gets the fix that worked last time.
- pgvector
- Full-text BM25
- Reciprocal rank fusion
Agents and orchestration
Agents watch sensor and quality streams on a schedule and raise a work order when a pattern matches a known failure. They do not touch controls. The proposal states what it saw and what it recommends.
- agent-worker
- Scheduled runs
- Proposal queue
Evaluation
Failure prediction is backtested against your own downtime history, so precision and recall are known numbers before anyone reorganizes a maintenance schedule around it. False alarms have a cost and we measure it.
- Backtesting
- Precision and recall
- quality-worker
Models
Classic predictive ML on sensor and downtime data, because anomaly detection and remaining-useful-life are tabular problems with decades of method behind them. Vision models for inspection. Frontier models only for the language work.
- Anomaly detection
- Vision inspection models
- Frontier models, one gateway, routed per task
Data boundary
Process data, yields and recipes are trade secrets. They stay in your accounts, scoped per line and per plant, under no-training API terms.
- Supabase row-level security
- Per-site scoping
- No-training API terms
MES and ERP, Sensor history, Maintenance logs feed a hybrid index. The agent runtime works from that index, and every consequential action passes a human approval before it reaches Ask your brain, Workflow builder, Cost per outcome.
Your systems
- 01MES and ERP
- 02Sensor history
- 03Maintenance logs
The system
- 01Hybrid index
- 02Agent runtime
- 03Your approvalhuman
Where your team works
- 01Ask your brain
- 02Workflow builder
- 03Cost per outcome
MES and ERP, Sensor history, Maintenance logs and the system run inside a boundary labeled your accounts. The only path that crosses the boundary is the audited egress to the model API, under no-training terms. The boundary is what answers Customer quality standards (ISO 9001, AS9100, IATF 16949) and Safe Food for Canadians Regulations and HACCP.
Inside your accounts
- 01MES and ERP
- 02Sensor history
- 03Maintenance logs
- 04System in your cloud
- 05Audit log
Outside, through the audited port
- 01Model API
How we build it for manufacturers
Capture what the two people who know actually know.
The first build turns tribal troubleshooting knowledge into an indexed, searchable record. It pays for itself the first time someone retires or calls in sick.
Backtest failure prediction before changing a schedule.
Precision and recall are measured against your own downtime history. False alarms have a cost and we report it separately rather than as one accuracy number.
Add vision inspection where volume justifies it.
Bespoke vision is justified by throughput and latency, not novelty. Below that bar an on-demand model is cheaper and easier to maintain.
What we will not automate
Anything that touches controls or operational technology. The system raises a work order and a person acts on it.
What that means in practice
Classic and predictive ML
Not every problem needs a language model. Predicting numbers, churn, demand, fraud risk, is usually solved better, cheaper and more explainably with proven statistical ML.
Where we stop. When the input is language, judgment or unstructured documents, classic ML underperforms and we say so. The discipline runs both ways: if your problem is a prediction problem, you will hear that it does not need an LLM from us before you pay for one.
Vision, documents and speech
AI for eyes and ears: reading documents, watching camera feeds, transcribing calls.
Where we stop. Bespoke computer vision is justified by volume and latency, not by novelty. Below that bar, a vision-language model on demand is cheaper to run and easier to maintain, and we will tell you which side of the bar you are on before anything is built.
MLOps and infrastructure
The plumbing that keeps AI running in production: deployment, scaling, monitoring and retraining, so it does not quietly degrade after launch.
Where we stop. We do not park your system on infrastructure only we understand. If your platform team runs Kubernetes on AWS, we deploy there. The architecture has to survive us leaving; that is the point of it.
Where your team works
Tour the platformAsk your brain
Ask a question. Get the answer and where it came from.
Every answer cites the document behind it, so you can check the source yourself. When the brain does not have the answer, it says so instead of guessing.
Workflow builder
Describe the workflow. Watch it assemble.
Say what should happen in plain language and the builder assembles the automation on a canvas you can read, run, and change.
Cost per outcome
Every dollar of spend traced to the work behind it.
Outcomes delivered, cost per outcome, value attributed, return on spend. The same measurement the guarantee is settled against, live on one page.
Task board
People and agents, working the same board.
Every piece of work lives on one shared board, whether a person or an agent owns it. Handoffs between the two are explicit, so nothing falls in the gap.
By team
The same system, seen from the desk that runs it.
The hard questions
Free · 3-5 days
Know your number in five days.
We map your operations, find the highest-ROI automations, and hand you a ranked plan with the payback math attached. Yours to keep, whoever builds it.
Prefer to talk first? Book 15 minutes with James. No pitch deck.