labs@atheron:~/industries/manufacturing/ai$ train agent --grounded

AI for manufacturing

Vision quality checks and maintenance agents on your data.

In a plant, AI earns its place in two jobs: seeing defects a tired person misses, and finding the answer in a thousand pages of manuals. Both work best on models trained or grounded on your own data, often running on your own hardware.

Start a project See the example projects

// what matters here

What is different in this industry.

Your images never leave the plant

A vision model can run on a GPU server in the plant. We price buying one against renting, and the estimate shows when the purchase pays for itself.

Measured before it goes live

Every model is scored against a held-out set of your own labelled images or questions, and the score is a gate: it does not ship below the agreed line.

People stay in the loop

The model flags; a person confirms. Overrules are recorded and become the next round of training data.

Answers with their source

A maintenance assistant cites the manual page or the work order behind every answer, so a technician can check it before acting on it.

// example projects

Priced examples for this service.

Each one is hypothetical, labelled as such, and priced live from our rate card. Open any of them in the estimator and make it yours.

Example projectWeb appAICustom LLM

Vision quality check on the line

Hypothetical. Not a client, not a result.

// The problem

Surface defects are caught by a person at the end of the line, who misses some and cannot keep up at full speed.

// What we would build

A custom vision model, fine-tuned on the plant's own images, running on a GPU server in the plant so images never leave the site, with a review screen for quality staff.

// What is in it

Image capture at the station, labelled by your quality team in a simple tool
A model fine-tuned on your defects, measured against a held-out set before it goes live
Inference on a GPU server on site, with no images sent to the cloud
A review queue where staff confirm or overrule what the model flagged
Defect rates by line, shift and supplier lot

// Stack

PyTorch
ONNX Runtime
Python
Next.js
Lenovo GPU server on site

// estimate

Build: ≈ US$164,000 to US$249,000, delivered within 40 weeksCAD 233,400 to 355,000
Hosting: ≈ US$176 a monthCAD 250 a month
Support: ≈ US$2,500 a monthCAD 3,565 a month
Hardware: ≈ US$47,800 to US$197,000CAD 68,000 to 280,800

Prices in your currency are estimates from today's Bank of Canada rate. All invoicing is in CAD or USD.

AI route: Custom LLM
Model running cost: ≈ US$176 a monthCAD 250 a month

Timeline by milestone

Discovery: 2.9 to 3.4 weeks
Data and evaluation plan: 0.4 weeks
Design approved: 2.4 to 3.9 weeks
Core features: 1.9 to 3.4 weeks
Full build: 1.4 to 2.4 weeks
Training: 8.9 to 15.4 weeks
Testing and fixes: 0.4 weeks
Evaluation and red-teaming: 3.9 to 4.4 weeks
Traceability check passed: 3.9 to 4.4 weeks
Launch: 0.9 to 1.4 weeks
Production: 0.9 to 1.4 weeks

Outside our hands, and added to the calendar

Hardware delivery, after it is ordered: 2 to 6 weeks

How it is paid

Deposit 20%: CAD 46,440 to 70,760
Discovery 1.1%: CAD 2,554.20 to 3,891.80
Data and evaluation plan 7.2%: CAD 16,718.40 to 25,473.60
Design approved 1.1%: CAD 2,554.20 to 3,891.80
Core features 3.5%: CAD 8,127 to 12,383
Full build 2.3%: CAD 5,340.60 to 8,137.40
Training 21.7%: CAD 50,387.40 to 76,774.60
Testing and fixes 1.1%: CAD 2,554.20 to 3,891.80
Evaluation and red-teaming 14.4%: CAD 33,436.80 to 50,947.20
Traceability check passed 8.7%: CAD 20,201.40 to 30,780.60
Launch 1.1%: CAD 2,554.20 to 3,891.80
Production 7.8%: CAD 18,111.60 to 27,596.40
Holdback, 30 days after launch (10%): CAD 23,220 to 35,380
GPU time for training and testing, at cost: CAD 1,200. Billed up front, at cost, outside the milestones
GPU server for AI, for the model: CAD 68,000 to 280,800. Billed up front, at cost, outside the milestones

Start from this example

Example projectWeb appAIOpen-source model

Maintenance manuals assistant

Hypothetical. Not a client, not a result.

// The problem

Technicians search through hundreds of PDF manuals and old work orders to find how a fault was fixed last time.

// What we would build

An assistant that answers from the plant's own manuals and work order history, with the page it came from, built on an open-source model hosted in your own cloud account so the documents stay yours.

// What is in it

Manuals, drawings and past work orders indexed and kept in sync
Answers that cite the manual page or the work order they came from
Tools to look up a machine's history and open a work order
Evaluation set from real questions, run before every change
An open-source model served on rented GPUs you control

// Stack

Llama or Qwen (open-source)
vLLM
PostgreSQL with pgvector
Next.js
DigitalOcean GPU

// estimate

Build: ≈ US$96,100 to US$147,000, delivered within 22 weeksCAD 136,800 to 209,100
Hosting: ≈ US$4,360 a monthCAD 6,210 a month
Support: ≈ US$1,000 a monthCAD 1,425 a month

Prices in your currency are estimates from today's Bank of Canada rate. All invoicing is in CAD or USD.

AI route: Open-source model
Model running cost: ≈ US$4,190 a monthCAD 5,960 a month

Timeline by milestone

Discovery: 2.9 to 3.4 weeks
Specification and evaluation plan: 0.4 weeks
Design approved: 2.4 to 3.4 weeks
Core features: 1.4 to 1.9 weeks
Full build: 0.9 to 1.4 weeks
Working pilot: 3.9 to 6.9 weeks
Testing and fixes: 0.9 to 1.9 weeks
Launch: 0.9 to 1.4 weeks
Production: 0.9 to 1.4 weeks

How it is paid

Deposit 20%: CAD 27,300 to 41,760
Discovery 1.8%: CAD 2,457.00 to 3,758.40
Specification and evaluation plan 10.5%: CAD 14,332.50 to 21,924.00
Design approved 1.8%: CAD 2,457.00 to 3,758.40
Core features 5.6%: CAD 7,644.00 to 11,692.80
Full build 3.7%: CAD 5,050.50 to 7,725.60
Working pilot 31.7%: CAD 43,270.50 to 66,189.60
Testing and fixes 1.8%: CAD 2,457.00 to 3,758.40
Launch 1.8%: CAD 2,457.00 to 3,758.40
Production 11.3%: CAD 15,424.50 to 23,594.40
Holdback, 30 days after launch (10%): CAD 13,650 to 20,880
GPU time for training and testing, at cost: CAD 300. Billed up front, at cost, outside the milestones

Start from this example

// questions

Questions we get about this.

How many images do we need to train a defect model?

It depends on how varied the defects are. Fine-tuning an existing vision model often starts from a few hundred labelled examples per defect type; we measure as we go and tell you when more data would help.

Frontier, open-source or custom?

For vision on the line, a fine-tuned custom model on site is usually right. For a manuals assistant, an open-source model in your cloud keeps the documents private; a frontier model is cheaper to start if the manuals are not sensitive. Each route is priced on its own.

What does it cost to run each month?

The estimate shows it separately from the build: GPU rental or the server you bought, plus hosting. Open an example in the estimator to see both.

Back to the industry: Manufacturing About the service: AI agents and models

// next

Not quite your project?

Tell us what you have in mind. We will come back to you with a range and the questions that would narrow it. Or book a call and talk it through.

Start a project Book a call