A predictable process for an unpredictable technology.
- 011-2 weeks
Discovery
Workshops, data audit, and technical feasibility assessment to define the roadmap.
- 021-2 weeks
Architecture
System design, model selection, and the evaluation framework we will hold the build to.
- 034-12 weeks
Build
Sprints with weekly demos. Production infrastructure from the first commit.
- 041-2 weeks
Launch
Staged rollout, load testing, A/B testing, and handoff documentation.
- 05Ongoing
Operate
MLOps, continuous evaluation, prompt optimisation, and model upgrades.
What happens in each stage
The durations are planning ranges. Discovery ends with a roadmap for your project, and what you learn there sets the real timeline.
Discovery
1-2 weeks
What happens
- Workshops with the people who own the problem.
- An audit of the data the system would use.
- A trial of the task on a strong base model, using real examples, to see where it fails. The failures tell us what the project needs.
What you receive
A written feasibility assessment and a roadmap. If building is the wrong answer, the assessment says so.
What we need from you
Sample data, and one person who can tell a correct answer from a wrong one.
Architecture
1-2 weeks
What happens
- We pick the approach that matches the failures found in discovery: better prompting, retrieval, fine-tuning, or an agent that uses tools.
- We select models and design the system around your infrastructure and data rules.
- We write the evaluation framework: real test cases, a scoring method, and the score that counts as done.
What you receive
A system design, the reasons behind the model choice, and the evaluation set the build will be held to.
What we need from you
Agreement on what a correct answer looks like, and a review of the test cases.
Read more: RAG vs fine-tuning, How to evaluate a RAG system
Build
4-12 weeks
What happens
- Sprints that end in a weekly demo.
- Production infrastructure from the first commit, so nothing has to be rebuilt to launch.
- Every change is scored against the evaluation set, so a fix that breaks something else shows up the same day.
What you receive
Working software at each demo, running in your environment where your rules require it.
What we need from you
Someone at each demo who can react to real outputs, and quick answers on edge cases.
Launch
1-2 weeks
What happens
- A staged rollout to a small group of users first.
- Load testing and A/B testing before wider release.
- Handoff documentation for the people who will run the system.
What you receive
The system in production, with monitoring in place and documentation your team can use.
What we need from you
A named person who makes the go or no-go call, and a first group of real users.
Operate
Ongoing
What happens
- MLOps: deployment, monitoring, and cost tracking.
- Continuous evaluation. Every model upgrade or prompt change is scored against the same evaluation set.
- Failures found in production become new test cases.
What you receive
A system that is measured after launch as well as before it.
What we need from you
Access to user feedback and flagged answers.
Ready to move from AI experiments to AI in production?
A thirty-minute call. We assess the use case, name the risks, and tell you honestly whether we are the right fit.