MAP · Manufacturing Agentic Platform
One model of your operation.
Agents that ask before they act.
MAP connects the systems you run, reconciles them into one shared model, and runs agents on it under human control.
reading machine status, ERP sync 08:10, the LASER-01 queue
Illustrative example. Jobs, times and figures are made up.
One disruption, walked through MAP.
Scroll through it. The floor follows each step: what MAP reads, what the agents propose, and where a person decides.
-
01 · Connect
It reads what you already run.
ERP, quality records, engineering files and the spreadsheets that hold the real schedule. Read-only, through one supported export or API, mapped field by field with someone who knows your shop.
-
02 · One model
The floor gets a shared model.
Jobs, routings, work centers, parts and quotes, reconciled across systems. Every product and every agent works from that one model, so a scheduler and an estimator never disagree about what's on the floor.
-
03 · Disruption
07:29. LASER-01 goes down.
Six jobs are waiting in its queue. A detector written in code opens one incident with one owner. A language model never decides whether something happened.
-
04 · Forks
Agents propose. The solver checks.
Each option is a fork of the plan, never an edit to it. Moving all six jobs fails a rule: WO-4471 needs a bigger bed than LASER-02 has. Moving five keeps every date for 3.5 hours of overtime.
-
05 · Approval
The planner decides.
Only an approval publishes the fork as the new plan. Releasing work, shipping, signing and promising a date to a customer stay with people at every stage.
-
06 · Record
The decision goes on the record.
What it relied on, the forks it weighed, who decided and when. When the jobs ship, the promise gets scored against the real date, and you see the score.
Integration is a mapping someone approves. Not a coding project.
Every ERP names things its own way. A job is a Job in Epicor, a Production Order in Business Central and a Manufacturing Order in Odoo. And "complete" can mean done, shipped or closed in the books.
So MAP doesn't guess. It proposes, and a person who knows your shop confirms.
- Code
Profile
Read the source structure and sample values.
- Agent
Propose
Suggest how each field maps to the shared model, with the evidence for it.
- Our integration lead
Preview
Show each mapping against your real values before anything imports.
- Your data owner
Approve
Someone who knows the shop confirms what each field means.
- Code + integration lead
Reconcile
Match identities and timestamps, find the gaps, and flag any match we're unsure of instead of merging it.
- Code
Monitor
Track freshness, match errors and drift for as long as it runs.
Read-only first. Every mapping is versioned. Money stays in your ERP and accounting system.
One model of the operation, and an owner for every fact.
The model is versioned. Every table, tool and glossary term in MAP is generated from it, so nothing downstream drifts out of step.
Your systems keep their own records. The model only says which system is the source of truth for which fact.
| ERP | Customers, quotes, orders, routings, work orders, material, shipments, costs |
|---|---|
| PLM / PDM | Parts, revisions, bills of materials, change orders, drawings |
| QMS | Inspections, nonconformances, corrective actions, gauges, first articles |
| Floor | Operation runs, check-ins, downtime |
| MAP | Schedules, approved shop rules, and the record of every proposal, decision and outcome |
Autonomy is earned one rung at a time.
Each kind of action starts at the bottom. It climbs only on your plant's own record, and only when a named person grants the next rung. One reverted action drops it back down. Pilots run on the first two.
- Rung 0
Observe
Read and explain.
PilotsWhy is this job late?
- Rung 1
Suggest
Propose a fork for a person to review.
PilotsFour recovery options for a laser outage.
- Rung 2
Safe act
Act, but only in the safe direction.
Place a hold. Flag a lot.
- Rung 3
Act with notice
Carry out an approved kind of action within set limits, then report it.
Re-sequence inside one cell.
When your planner overrides a proposal, that's the lesson.
A correction becomes a rule only after it's been classified, tested on past cases and approved. Rollback is always available.
- 01
Override
Someone changes a proposal and says why.
- 02
Classify
Missing rule, one-off exception, commercial choice, data defect or plain error.
- 03
Propose
A candidate rule, scoped to your shop.
- 04
Test
Replayed against past cases before anyone relies on it.
- 05
Approve
A named person switches it on.
Your rules improve your proposals and nobody else's. Knowledge from one customer never moves to another.
Built to trust the model the least.
Models do the reading and the explaining. Code, policies and named people hold the authority, and nothing a model says can switch them off.
Approval for anything that leaves
Messages to customers or vendors, production schedule changes and any ERP write-back all wait for a person.
Documents are data
An RFQ, an email or a drawing can't instruct an agent, whatever text is hidden in it.
Allowlisted tools
Agents get scoped permissions and named tools. Never open database or shell access.
Numbers are traced
Every number and ID in an answer has to match a tool result. Dates come as ranges. "I don't know" is allowed, and it names the missing data.
Visible when it fails
If a sync, a model or the solver fails, you get a labeled read-only or stale-data view, not a guess.
Private models, by tier
Agents ask for a tier, not a vendor, so models stay swappable. First deployments run self-hosted, inside your boundary.
Configured for you. Never forked for you.
Every customer runs the same MAP, set up for how they work. Each engagement fits one of three shapes, and we tell you which before any work starts.
Included
Configuration
Most of what we deliver. Your routings, calendars, priorities, approval rules, report formats and units, set up in software we already maintain.
Priced as implementation
A reusable extension
When you need something we haven't built, we build it into MAP so the next customer can use it too.
Quoted separately
One-off work
Work only one customer will ever need. We quote it on its own, and turn down work that would split the product in two.
What we're building toward on MAP.
None of this is available yet. We build each piece when paid use asks for it, and we'll tell you which parts are live before you pay for anything.
One companion per person
One voice for the whole crew, set to that person's role, permissions and interruption budget. Behind it, shared specialists do the scheduling, estimating and quality work, and sign their parts of each answer.
Ask the floor
Pick a machine or a job and ask why. Station answers come from live data, in code, with no model needed.
A brief before the shift
A short morning brief where every line ends in a decision. What the crew already handled goes in the digest.
Where people already work
The app first, then phone, email and SMS, then Microsoft Teams. A chat button can only approve what can be undone.
The quote meets the schedule
A configurator quote whose lead time is checked against real capacity in Cadence.
What technical buyers ask first.
No. You buy Cadence, Company Brain or a configurator, and MAP comes with whichever product you start with. A second product starts from the model the first one already built.
Each pilot starts with one supported API or controlled export, chosen after we've looked at your data. We add systems as customers bring them, and we don't claim to connect to everything.
Not during a pilot. We consider write-back only after the mappings are versioned and reconciliation has been tested, and it always needs approval. Money stays in your ERP and accounting system; agents never post to the books.
You see a clearly labeled read-only or stale-data state. A failed sync, model or solver run never turns into a quiet guess.
Agents ask for a tier, never a vendor, and each deployment maps the tier to a model. We choose models by testing them on our own tasks: pulling fields from documents, answering from sources, refusing when evidence is missing, speed and cost. First deployments run them self-hosted.
Want the technical walkthrough?
or write to info@quogen.ai
Bring your IT or data lead. We'll go through mappings, hosting and approvals with them.
We reply to every message ourselves. If we're not the right fit, we'll say so, and usually suggest who is.