Skip to content
AtheronLABS

You're visiting from the United States. Prices are shown in US dollars. Not right?

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

Transit AI

Rider feedback sorted and questions answered.

Transit agencies get more rider comments and questions than staff can read. A model trained on your own past feedback sorts them by route and topic, and an assistant that answers from your help articles takes the routine questions, so staff spend time on the ones that need a person.

// what matters here

What is different in this industry.

01

Trained on your feedback

A small open model fine-tuned on comments your staff labelled, served in your own cloud account.

02

Corrected by staff

Staff fix tags in a review screen, and the corrections feed the next training round.

03

Answers that cite

The help assistant answers from your articles and schedules and names the article it used.

04

A person when needed

The assistant hands over to staff with the conversation attached when it cannot help.

// 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

Rider feedback sorting

Hypothetical. Not a client, not a result.

// The problem

Thousands of rider comments and complaints a month arrive by form, email and social media, and staff sort them by hand into routes, stops and topics weeks after the fact.

// What we would build

A small open model fine-tuned on the agency's own labelled feedback, served in its own cloud account, that tags each comment with route, stop, topic and urgency, with a screen for staff to correct the tags so the model keeps improving. Rider comments never leave your account.

// What is in it

  • A model fine-tuned on your past feedback, labelled with your staff
  • Every comment tagged by route, stop, topic and urgency
  • A review screen where staff correct tags, feeding the next training round
  • Trends by route and topic for service planners
  • An evaluation set run before every new version goes live

// Stack

  • Llama or Mistral (fine-tuned)
  • vLLM
  • PostgreSQL
  • Next.js
  • AWS GPU

// estimate

Build
≈ US$137,000 to US$209,000, delivered within 31 weeksCAD 194,600 to 297,000
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

Your IT security review and sign-on setup
4 to 8 weeks

How it is paid

Deposit 20%
CAD 38,680 to 59,160
Discovery 1.5%
CAD 2,901 to 4,437
Data and evaluation plan 7.9%
CAD 15,278.60 to 23,368.20
Design approved 1.5%
CAD 2,901 to 4,437
Core features 4.6%
CAD 8,896.40 to 13,606.80
Full build 3.1%
CAD 5,995.40 to 9,169.80
Training 23.9%
CAD 46,222.60 to 70,696.20
Testing and fixes 1.5%
CAD 2,901 to 4,437
Evaluation and red-teaming 15.9%
CAD 30,750.60 to 47,032.20
Launch 1.5%
CAD 2,901 to 4,437
Production 8.6%
CAD 16,632.40 to 25,438.80
Holdback, 30 days after launch (10%)
CAD 19,340 to 29,580
GPU time for training and testing, at cost
CAD 1,200. Billed up front, at cost, outside the milestones
Example projectWeb appAIFrontier model

Multilingual rider help centre

Hypothetical. Not a client, not a result.

// The problem

A city's transit and parking questions arrive by phone and email in many languages, and the answers live in a website few riders can navigate.

// What we would build

A help centre in English, French and the languages your riders speak, with an assistant on a frontier model that answers from the help articles and your published schedules and fare rules, cites the article, and hands over to a person when it cannot help.

// What is in it

  • Help articles on fares, passes, accessibility, parking and lost property
  • An assistant that answers from your articles and published GTFS schedules
  • Every answer cites the article it comes from
  • A hand-over to staff with the conversation attached
  • Articles translated and reviewed by native speakers

// Stack

  • Next.js
  • Frontier model API
  • PostgreSQL with pgvector
  • GTFS
  • DigitalOcean Toronto

// estimate

Build
≈ US$91,900 to US$141,000, delivered within 22 weeksCAD 130,900 to 200,100
Hosting
≈ US$558 a monthCAD 795 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
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 3.9 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

How it is paid

Deposit 20%
CAD 26,180 to 40,020
Discovery 2%
CAD 2,618 to 4,002
Specification and evaluation plan 10.2%
CAD 13,351.80 to 20,410.20
Design approved 2%
CAD 2,618 to 4,002
Core features 6.2%
CAD 8,115.80 to 12,406.20
Full build 4.1%
CAD 5,366.90 to 8,204.10
Working pilot 30.7%
CAD 40,186.30 to 61,430.70
Testing and fixes 2%
CAD 2,618 to 4,002
Launch 2%
CAD 2,618 to 4,002
Production 10.8%
CAD 14,137.20 to 21,610.80
Holdback, 30 days after launch (10%)
CAD 13,090 to 20,010

// questions

Questions we get about this.

Why train our own model to sort rider feedback?

Your routes, stops and topics are your own, and rider comments stay in your account. A frontier model costs less to start if you are comfortable sending comments to a provider.

Can the transit assistant promise a refund or a fare exception?

No. It answers from your published rules and hands anything else to staff.

How do we know the feedback model is right?

An evaluation set of comments your staff labelled is run before every new version goes live.

// 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.

AI for transit agencies: rider feedback sorting and help assistants | Atheron Network Labs