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

AI for banks

AML alert triage and branch procedures assistants.

A bank cannot send customer transactions to someone else's model, and model risk management expects every model to be explained, validated and monitored. So banking AI here is custom or open-source, on hardware or accounts the bank controls, with the validation evidence produced as part of the build.

Start a project See the example projects

// what matters here

What is different in this industry.

Customer data stays in the bank

The AML model runs on GPU servers in your data centre and the procedures assistant in your own cloud account. Nothing is sent to a model provider.

Evidence for model validation

We document the training data, the method, the held-out results and the limits of each model, in the form your model risk team needs to review it.

Analysts decide, the model ranks

The triage model orders the queue and shows its reasons. Decisions and any report to FINTRAC stay with your analysts.

Monitored after launch

The model's scores are compared with analyst outcomes on a schedule, and drift raises an alert before it becomes a finding.

// 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

AML alert triage on your servers

Hypothetical. Not a client, not a result.

// The problem

Transaction monitoring produces a queue of alerts analysts cannot clear, and true suspicious activity waits behind thousands of false positives.

// What we would build

A custom model fine-tuned on the bank's own closed alerts and analyst notes, running on GPU servers in the bank's data centre so customer transactions never leave it, that ranks each alert and explains why for the analyst who decides.

// What is in it

A model fine-tuned on your past alerts and how analysts closed them
Each alert ranked with the transactions and patterns behind the score
The analyst decides and writes the case; the model never files a report
Measured against a held-out set of closed alerts before it goes live, and on a schedule after
Runs on servers in your data centre, with no data sent outside

// Stack

PyTorch
Python
PostgreSQL
Next.js
Lenovo GPU servers on premises

// estimate

Build: ≈ US$193,000 to US$294,000, delivered within 38 weeksCAD 274,600 to 418,100
Hosting: ≈ US$176 a monthCAD 250 a month
Support: ≈ US$5,010 a monthCAD 7,135 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: 2.9 to 4.9 weeks
Core features: 1.9 to 3.4 weeks
Full build: 1.4 to 2.4 weeks
Training: 7.4 to 12.4 weeks
Testing and fixes: 0.4 weeks
Evaluation and red-teaming: 3.9 to 4.4 weeks
Security review 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
Core sandbox and credentials from your provider: 1 to 4 weeks
Penetration test by your chosen firm: 2 to 4 weeks

How it is paid

Deposit 20%: CAD 54,680 to 83,380
Discovery 1.3%: CAD 3,554.20 to 5,419.70
Data and evaluation plan 6.9%: CAD 18,864.60 to 28,766.10
Design approved 1.3%: CAD 3,554.20 to 5,419.70
Core features 4%: CAD 10,936 to 16,676
Full build 2.7%: CAD 7,381.80 to 11,256.30
Training 20.9%: CAD 57,140.60 to 87,132.10
Testing and fixes 1.3%: CAD 3,554.20 to 5,419.70
Evaluation and red-teaming 13.9%: CAD 38,002.60 to 57,949.10
Security review passed 8.7%: CAD 23,785.80 to 36,270.30
Launch 1.3%: CAD 3,554.20 to 5,419.70
Production 7.7%: CAD 21,051.80 to 32,101.30
Holdback, 30 days after launch (10%): CAD 27,340 to 41,690
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

Assistant for branch staff

Hypothetical. Not a client, not a result.

// The problem

Front-line staff search a shared drive of procedures, bulletins and rate sheets to answer a member's question, and often find an out-of-date version.

// What we would build

An assistant that answers from the current procedures only and cites the document and section, on an open-source model served in the bank's own AWS account in Canada so internal policy stays inside it.

// What is in it

Procedures, bulletins and forms indexed, with retired versions excluded
Answers that cite the document, section and effective date
Owners of each procedure see the questions asked about it
An evaluation set of real branch questions, run before every update
An open-source model on GPUs in your own cloud account

// Stack

Llama or Mistral (open-source)
vLLM
PostgreSQL with pgvector
Next.js
AWS Canada

// estimate

Build: ≈ US$95,400 to US$146,000, delivered within 19 weeksCAD 135,900 to 207,700
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: Open-source model
Model running cost: ≈ US$6,530 a monthCAD 9,300 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: 2.9 to 4.9 weeks
Testing and fixes: 0.9 to 1.4 weeks
Launch: 0.9 to 1.4 weeks
Production: 0.9 to 1.4 weeks

How it is paid

Deposit 20%: CAD 27,120 to 41,480
Discovery 2.3%: CAD 3,118.80 to 4,770.20
Specification and evaluation plan 9.6%: CAD 13,017.60 to 19,910.40
Design approved 2.3%: CAD 3,118.80 to 4,770.20
Core features 7.1%: CAD 9,627.60 to 14,725.40
Full build 4.7%: CAD 6,373.20 to 9,747.80
Working pilot 29%: CAD 39,324 to 60,146
Testing and fixes 2.3%: CAD 3,118.80 to 4,770.20
Launch 2.3%: CAD 3,118.80 to 4,770.20
Production 10.4%: CAD 14,102.40 to 21,569.60
Holdback, 30 days after launch (10%): CAD 13,560 to 20,740
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.

Why a custom model for AML instead of a frontier one?

Two reasons: customer transactions should not leave the bank, and your alert history holds patterns specific to your book. A model fine-tuned on that history, run in your data centre, answers both.

Will the AML triage model close alerts on its own?

No. It ranks and explains; an analyst works every alert that policy says must be worked. You decide whether low-ranked alerts are handled differently, under your own policy.

How does the assistant avoid citing a retired policy?

Each document carries its effective and retirement dates, retired versions are removed from the index, and every answer shows the date of the source so staff can see it is current.

Back to the industry: Banking 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