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

AI for drug safety

Safety cases and SOPs, confirmed by a specialist.

Life sciences runs on reading: safety reports, SOPs, batch records. AI helps when it reads in the domain's own language and shows its source, and when a qualified person still makes the call. We build on custom or open-source models that stay inside your environment.

Start a project See the example projects

// what matters here

What is different in this industry.

Speaks the domain's language

A safety model trained further on medical and pharmacovigilance text knows MedDRA terms and drug names that a general model guesses at.

A specialist confirms

The model proposes case fields and terms; a safety specialist confirms or corrects them. Corrections are logged and become the next round of training data.

Effective versions only

An SOP assistant reads from the effective version and cites document, version and section, so nobody acts on a superseded procedure.

Validated like any GxP tool

Each model is scored on a held-out set of your own cases, and the score, data and model version are recorded so a change is controlled.

// 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

Safety case intake model

Hypothetical. Not a client, not a result.

// The problem

A pharmacovigilance team reads every incoming adverse event report by hand to find the product, the event, the patient details and the reporting clock.

// What we would build

A custom model, further trained on medical and safety literature and then on the team's own de-identified cases, that reads each report and proposes the case fields and MedDRA terms, which a safety specialist then checks, running on a GPU server inside the company.

// What is in it

Reports from email, web forms and call notes read into case fields
Suggested MedDRA terms with the passage they came from
Seriousness and reporting deadline flagged for a specialist to check
A model trained further on safety language, then on your cases
Scored against a held-out set of your cases before every release

// Stack

Open-weight base model, further trained
PyTorch
Python
Next.js
Lenovo GPU server on site

// estimate

Build: ≈ US$269,000 to US$407,000, delivered within 55 weeksCAD 382,400 to 579,600
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: 2.9 to 4.9 weeks
Core features: 2.4 to 3.9 weeks
Full build: 1.4 to 2.9 weeks
Training: 16.4 to 28.4 weeks
Testing and fixes: 0.4 weeks
Evaluation and red-teaming: 3.9 to 4.4 weeks
Validation package delivered: 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
Vendor test environment access: 1 to 4 weeks

How it is paid

Deposit 20%: CAD 74,980 to 114,420
Discovery 0.7%: CAD 2,624.30 to 4,004.70
Data and evaluation plan 6.8%: CAD 25,493.20 to 38,902.80
Design approved 0.7%: CAD 2,624.30 to 4,004.70
Core features 2.2%: CAD 8,247.80 to 12,586.20
Full build 1.4%: CAD 5,248.60 to 8,009.40
Training 20.4%: CAD 76,479.60 to 116,708.40
Testing and fixes 0.7%: CAD 2,624.30 to 4,004.70
Evaluation and red-teaming 13.6%: CAD 50,986.40 to 77,805.60
Validation package delivered 15.5%: CAD 58,109.50 to 88,675.50
Launch 0.7%: CAD 2,624.30 to 4,004.70
Production 7.3%: CAD 27,367.70 to 41,763.30
Holdback, 30 days after launch (10%): CAD 37,490 to 57,210
GPU time for training and testing, at cost: CAD 7,500. 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

SOP and documents assistant

Hypothetical. Not a client, not a result.

// The problem

Quality and manufacturing staff search a document system for the effective version of an SOP, and the section they need is buried in a long PDF.

// What we would build

An assistant that answers from effective SOPs only, cites the document, version and section, and opens a deviation form when asked, on an open-source model in the company's own cloud account so controlled documents stay private.

// What is in it

Effective SOPs, work instructions and forms indexed from your document system
Superseded versions excluded automatically
Every answer cites the document, version and section
A tool to start a deviation or change request with the context filled in
Evaluation set of real questions, run before every change

// Stack

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

// estimate

Build: ≈ US$130,000 to US$199,000, delivered within 24 weeksCAD 185,000 to 282,900
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 3.9 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: 0.4 to 0.9 weeks
Validation package delivered: 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

Vendor test environment access: 1 to 4 weeks

How it is paid

Deposit 20%: CAD 36,940 to 56,520
Discovery 1.4%: CAD 2,585.80 to 3,956.40
Specification and evaluation plan 8.2%: CAD 15,145.40 to 23,173.20
Design approved 1.4%: CAD 2,585.80 to 3,956.40
Core features 4.4%: CAD 8,126.80 to 12,434.40
Full build 2.9%: CAD 5,356.30 to 8,195.40
Working pilot 24.7%: CAD 45,620.90 to 69,802.20
Testing and fixes 1.4%: CAD 2,585.80 to 3,956.40
Validation package delivered 15.5%: CAD 28,628.50 to 43,803.00
Launch 1.4%: CAD 2,585.80 to 3,956.40
Production 8.7%: CAD 16,068.90 to 24,586.20
Holdback, 30 days after launch (10%): CAD 18,470 to 28,260
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 safety model, not a frontier API?

Safety reports carry personal health information and the language is specialised. A model trained further on the domain and run inside your company keeps the data in and reads the terms better. A frontier model is cheaper to start with for non-sensitive text, and we price both.

Can an AI tool be used in a GxP process?

It can, as a tool with a defined intended use, validated for that use and with a qualified person accountable for the result. We build it that way with your quality team.

How is an SOP assistant kept current?

It reads from your document system and re-indexes when a document becomes effective or is retired, so the answers follow your change control.

Back to the industry: Pharma and life sciences 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