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

AI for utilities

An operations assistant and feeder load forecasts.

Utilities have two good uses for AI that keep the data at home: finding the right procedure fast, and forecasting load as customers electrify. The first suits an open-source model inside your network, the second a custom model trained on your own history. Neither touches operational control.

Start a project See the example projects

// what matters here

What is different in this industry.

Inside your network

The assistant runs on a GPU server you own with no internet path, which suits the security rules utilities work under, such as NERC CIP.

Only current procedures

Superseded revisions are excluded from answers, and every answer names the document, revision and section.

Forecasts with honest ranges

The load model gives a range for each feeder, so planners see the uncertainty instead of a single confident line.

Advice, not control

Neither system operates equipment or changes settings. Operators and planners decide.

// 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 appAIOpen-source model

Control room assistant

Hypothetical. Not a client, not a result.

// The problem

Control room operators and field supervisors search binders and shared drives for the right switching procedure, operating instruction or manual, often in the middle of an event.

// What we would build

An assistant that answers from your own procedures, operating instructions and past switching orders, citing the document and section, on an open-source model running on a server inside your network so nothing touches the internet. It reads and cites; it never operates equipment.

// What is in it

Procedures, instructions and manuals indexed and kept in step with the document system
Answers that cite the document, revision and section
Superseded revisions excluded, so only current procedures are quoted
Evaluation set of real operator questions, run before every change
An open-source model on a GPU server inside your network
No connection to operational control systems

// Stack

Llama or Mistral (open-source)
vLLM
PostgreSQL with pgvector
Next.js
Lenovo GPU server on site

// estimate

Build: ≈ US$120,000 to US$183,000, delivered within 23 weeksCAD 170,900 to 259,900
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: Open-source model
Model running cost: ≈ US$176 a monthCAD 250 a month

Timeline by milestone

Discovery: 2.9 to 3.4 weeks
Specification and evaluation plan: 0.4 weeks
Design approved: 2.4 to 3.9 weeks
Core features: 1.4 to 2.4 weeks
Full build: 0.9 to 1.4 weeks
Working pilot: 4.4 to 7.4 weeks
Testing and fixes: 1.4 to 1.9 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
Read-only access approved by your OT team: 1 to 4 weeks
Penetration test by your chosen firm: 2 to 4 weeks

How it is paid

Deposit 20%: CAD 34,120 to 51,920
Discovery 1.8%: CAD 3,070.80 to 4,672.80
Specification and evaluation plan 10.5%: CAD 17,913 to 27,258
Design approved 1.8%: CAD 3,070.80 to 4,672.80
Core features 5.6%: CAD 9,553.60 to 14,537.60
Full build 3.7%: CAD 6,312.20 to 9,605.20
Working pilot 31.7%: CAD 54,080.20 to 82,293.20
Testing and fixes 1.8%: CAD 3,070.80 to 4,672.80
Launch 1.8%: CAD 3,070.80 to 4,672.80
Production 11.3%: CAD 19,277.80 to 29,334.80
Holdback, 30 days after launch (10%): CAD 17,060 to 25,960
GPU time for training and testing, at cost: CAD 300. 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 appAICustom LLM

Feeder load forecast model

Hypothetical. Not a client, not a result.

// The problem

A distribution planner forecasts load per feeder from last year's peaks, while heat pumps, EV chargers and rooftop solar change demand faster than the spreadsheet can follow.

// What we would build

A custom forecasting model, fine-tuned from an open time-series foundation model on your own meter, feeder and weather history, that forecasts load per feeder for the days and seasons ahead, with a web view for planners. A custom model fits because the patterns are yours and the data is sensitive; it is retrained as new meter data arrives.

// What is in it

Feeder, smart meter and weather history cleaned and joined
A time-series model fine-tuned on your data and measured against held-out periods
Forecasts per feeder with a range, not a single line
Heat pump, EV and solar adoption scenarios set by the planner
Retraining on a schedule as new data arrives

// Stack

PyTorch
Chronos or TimesFM (open time-series models)
Python
Next.js
PostgreSQL with TimescaleDB
AWS GPU

// estimate

Build: ≈ US$139,000 to US$213,000, delivered within 31 weeksCAD 198,400 to 302,700
Hosting: ≈ US$6,710 a monthCAD 9,550 a month
Support: ≈ US$2,500 a monthCAD 3,565 a month

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$6,530 a monthCAD 9,300 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 2.9 weeks
Full build: 1.4 to 1.9 weeks
Training: 6.4 to 11.4 weeks
Testing and fixes: 0.4 weeks
Evaluation and red-teaming: 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

Read-only access approved by your OT team: 1 to 4 weeks
Penetration test by your chosen firm: 2 to 4 weeks

How it is paid

Deposit 20%: CAD 39,440 to 60,300
Discovery 1.5%: CAD 2,958.00 to 4,522.50
Data and evaluation plan 7.9%: CAD 15,578.80 to 23,818.50
Design approved 1.5%: CAD 2,958.00 to 4,522.50
Core features 4.6%: CAD 9,071.20 to 13,869.00
Full build 3.1%: CAD 6,113.20 to 9,346.50
Training 23.9%: CAD 47,130.80 to 72,058.50
Testing and fixes 1.5%: CAD 2,958.00 to 4,522.50
Evaluation and red-teaming 15.9%: CAD 31,354.80 to 47,938.50
Launch 1.5%: CAD 2,958.00 to 4,522.50
Production 8.6%: CAD 16,959.20 to 25,929.00
Holdback, 30 days after launch (10%): CAD 19,720 to 30,150
GPU time for training and testing, at cost: CAD 1,200. Billed up front, at cost, outside the milestones

Start from this example

// questions

Questions we get about this.

Why an open-source model in the control room?

Because procedures and network data should not leave your network, and a model you host can run with no outside connection. A frontier model would be quicker to start but sends every question to the provider.

What data does a feeder load forecast need?

Feeder or substation load history, smart meter data where you have it, and weather. More history helps; we review what you have before quoting.

Can operators use the assistant in an emergency?

It can help find the right procedure quickly, but it is a reference tool. Your emergency procedures and operator judgement come first.

Back to the industry: Energy and utilities 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