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

AI for fintech

Support agents and underwriting document readers.

Fintech AI splits by what the model sees. Public help content can go to a frontier model and be live quickly. Bank statements and pay stubs should not leave your cloud, so they go to an open-source model you control. We pick the route by the data, then measure it.

Start a project See the example projects

// what matters here

What is different in this industry.

The data decides the route

Public material goes to a frontier model for the fastest start. Personal financial documents stay on an open-source model in your account, so no third party sees them.

A person makes the credit decision

The document reader extracts and flags; an analyst confirms. The model never approves or declines an application.

Every figure traced to its source

Each value the reader extracts is shown beside the line it came from, so checking it takes a glance rather than a search.

Measured on your own cases

We build an evaluation set from your real documents and questions, and every change to a prompt or a model is scored against it before release.

// 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 appAIFrontier model

Help centre support agent

Hypothetical. Not a client, not a result.

// The problem

Support volume rises with every marketing push, and most tickets ask questions the public help centre already answers.

// What we would build

A support agent on a frontier model that answers from the published help centre and hands anything about a specific account to a person, chosen because the source material is public and it is the fastest route to launch.

// What is in it

Answers grounded in the help centre, with a link to the article used
Hand-off to a person with the conversation attached, for anything account specific
No account data sent to the model provider
An evaluation set of real questions, run before every change
A dashboard of what people ask and where the help centre falls short

// Stack

Frontier model API
PostgreSQL with pgvector
Next.js
Zendesk or Intercom integration
DigitalOcean

// estimate

Build: ≈ US$88,600 to US$135,000, delivered within 21 weeksCAD 126,200 to 192,900
Hosting: ≈ US$558 a monthCAD 795 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: Frontier model
Model running cost: ≈ US$383 a monthCAD 545 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 6.4 weeks
Testing and fixes: 0.9 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

Readiness review with your auditor: 6 to 12 weeks

How it is paid

Deposit 20%: CAD 25,240 to 38,580
Discovery 2%: CAD 2,524 to 3,858
Specification and evaluation plan 10.2%: CAD 12,872.40 to 19,675.80
Design approved 2%: CAD 2,524 to 3,858
Core features 6.2%: CAD 7,824.40 to 11,959.80
Full build 4.1%: CAD 5,174.20 to 7,908.90
Working pilot 30.7%: CAD 38,743.40 to 59,220.30
Testing and fixes 2%: CAD 2,524 to 3,858
Launch 2%: CAD 2,524 to 3,858
Production 10.8%: CAD 13,629.60 to 20,833.20
Holdback, 30 days after launch (10%): CAD 12,620 to 19,290

Start from this example

Example projectWeb appAIOpen-source model

Underwriting document reader

Hypothetical. Not a client, not a result.

// The problem

Analysts retype income and expense figures from bank statements and pay stubs, which is slow and puts typing errors into credit decisions.

// What we would build

An agent on an open-source model, served in your own AWS account in Canada so financial documents never reach a third party, that reads each document, extracts the figures and shows the analyst every value beside the line it came from.

// What is in it

Bank statements, pay stubs and notices of assessment read into structured fields
Each figure shown beside the part of the page it was read from
The analyst confirms or corrects, and corrections are kept for the next evaluation
Flags for mismatched names, dates and totals across documents
An open-source model on GPUs in your own cloud account

// Stack

Qwen or Llama (open-source)
vLLM
Python
PostgreSQL
Next.js
AWS Canada

// estimate

Build: ≈ US$129,000 to US$198,000, delivered within 22 weeksCAD 184,300 to 281,700
Hosting: ≈ US$7,020 a monthCAD 10,000 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 4.4 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

Outside our hands, and added to the calendar

Sandbox and API access from the vendor: 1 to 4 weeks

How it is paid

Deposit 20%: CAD 36,800 to 56,280
Discovery 1.9%: CAD 3,496.00 to 5,346.60
Specification and evaluation plan 10.5%: CAD 19,320 to 29,547
Design approved 1.9%: CAD 3,496.00 to 5,346.60
Core features 5.7%: CAD 10,488.00 to 16,039.80
Full build 3.8%: CAD 6,992.00 to 10,693.20
Working pilot 31.6%: CAD 58,144.00 to 88,922.40
Testing and fixes 1.9%: CAD 3,496.00 to 5,346.60
Launch 1.9%: CAD 3,496.00 to 5,346.60
Production 10.8%: CAD 19,872.00 to 30,391.20
Holdback, 30 days after launch (10%): CAD 18,400 to 28,140
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.

Will the support agent give financial advice?

No. It answers from your help centre and is instructed to hand anything about advice, an account or a dispute to a person. We test those hand-offs as part of the evaluation set.

Is open source as good on bank statements?

For extracting figures from documents, current open-source models do well once they are evaluated on your formats. We measure on your documents before you commit, and say if a different route would serve better.

What runs each month after a fintech AI launch?

Model usage for a frontier route, or GPU hosting for an open-source one, plus the application hosting. The estimate shows both separately from the build.

Back to the industry: Finance and fintech 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