How we work

No surprises. Weekly demos. Fixed price ranges.

Most AI projects fail because the scope is vague, the data is worse than expected, and nobody finds out until week ten. Our five-phase process is designed to surface those problems early, when they're cheap to fix, and to ship working software to production in a predictable window.

The structure is the same whether we are building an AI agent, a retrieval system, or a full SaaS MVP. What changes is the depth of each phase, not the sequence. You always know which phase you are in, what comes out of it, and what it costs.

The five phases.

Discovery

1 week

What we do

Stakeholder interviews to map the actual problem
System and data audit. What you have, what quality it is
Success metric definition before we touch a model
Go / no-go recommendation based on what we find

What you get

A 2-page scope document and a written go / no-go recommendation. If the project is not viable, you find out now, not eight weeks in.

Architecture

3–5 days

What we do

Model selection based on your latency, cost, and accuracy requirements
Retrieval strategy design (chunking, embedding, re-ranking)
Tool call and agent loop design
Evaluation harness design, how we know it's working

What you get

A full technical architecture document. Every decision is written down with the trade-offs. Your engineers can read it and understand why we made each choice.

Build

3–8 weeks

What we do

Iterative one-week sprints against the architecture doc
Live working demo every Friday, not a status update
Async Slack updates throughout the week
Eval harness running continuously against each sprint

What you get

Working software every week. By the end of the sprint cycle you have a fully tested, production-ready system with eval scores you can actually read.

QA & Deployment

3–5 days

What we do

Load testing under realistic concurrency
Red-teaming for prompt injection and edge cases
Production deployment to your infrastructure
Monitoring dashboard and alert setup

What you get

A live system with a monitoring dashboard, a runbook, and a handoff meeting where we walk your team through everything.

Maintenance

Ongoing, optional

What we do

Weekly eval runs to catch model drift
Model updates as better options become available
Incident response within 4 business hours
Monthly performance report

What you get

A system that improves over time. Optional, month to month. You can end it any month, no lock-in.

What we commit to.

Fixed price range agreed before we start

We quote a price range after Discovery. We don't move the goalposts mid-project. If scope changes, we discuss it before any extra work starts.

Weekly working demos

Not status updates. Not slide decks. Every Friday you see the software running against real data. If something is off, you tell us Friday and we fix it Monday.

Full documentation and handoff

At close: architecture doc, runbook, deployment guide, and a walkthrough session. Your team can maintain the system without us, and we verify that before we call it done.

Why it works this way.

The single most expensive moment in an AI project is the one where everyone realises, in week ten, that the data was never good enough to support the idea. Discovery exists to move that moment to week one, where it costs a conversation instead of a budget. That is why we run a real data audit before quoting a fixed price, and why we are willing to recommend against a project we could have billed for.

The weekly demo is the other load-bearing habit. A status update describes progress; a demo proves it. When you watch the system run against real data every Friday, scope drift and quality problems surface while they are still small. There is no week-ten reveal because there is no week without a working build to look at.

Evaluation runs through the whole engagement rather than sitting at the end as a formality. We define what good looks like during Architecture, build the harness alongside the system, and measure against it every sprint. By the time you reach QA, the question is not whether the system works but whether the numbers clear the bar you approved, and you can read those numbers yourself.

Maintenance is optional on purpose. Models change, traffic changes, and an unwatched AI system drifts. We are happy to run that watch for you, but the handoff is built so your own team can do it instead. The measure of a good handoff is that you could fire us the next day and nothing would break.

Questions about how we work.

How long does a typical project take start to finish?: Most AI agent and RAG projects run 6–10 weeks from kickoff to production deployment, including Discovery, Architecture, Build, and QA phases. SaaS MVPs typically run 8–14 weeks. We give a firm timeline estimate at the end of the Discovery phase.
Can we skip the Discovery phase if we already know what we want?: We still run a condensed version. Discovery is where we audit your data, define success metrics, and confirm the architecture will work before we write production code. Skipping it adds risk to both of us. It costs one week and is priced into every engagement.
What if the project scope changes mid-build?: Changes happen. We discuss the impact on timeline and price before doing the extra work, never after. If a change is small, we absorb it. If it shifts the scope materially, we send a one-page addendum with the adjusted range for your approval.
What do you need from us during the project?: Less than most teams expect, but it is specific. We need access to the systems the work touches, a named decision-maker who can answer questions within a day or two, and an hour at the Friday demo. Discovery is the most input-heavy phase because that is where we learn your context. After that, the demos keep you in the loop without pulling your team off their own work.
Who owns the code and the models when the project ends?: You do, completely. Code is delivered into a repository you control, with IP transferred at project close. We build on model accounts and infrastructure in your name, so there are no credentials or licenses held hostage and no per-seat runtime fee to us. If you never speak to us again after handoff, the system keeps running.
Do you sign an NDA and work inside our security requirements?: Yes. We sign NDAs as a matter of course and work within your security and compliance constraints, including BAAs for healthcare data and deployment inside your own cloud tenancy where policy requires it. We raise those requirements during Discovery so the architecture is designed around them rather than retrofitted later.

Start with a discovery call.

Describe your project below. We'll read it, reply within one business day, and schedule a one-hour discovery session if it looks like a fit.

No surprises. Weekly demos. Fixed price ranges.

The five phases.

Discovery

1 week

What we do

Stakeholder interviews to map the actual problem
System and data audit. What you have, what quality it is
Success metric definition before we touch a model
Go / no-go recommendation based on what we find

What you get

A 2-page scope document and a written go / no-go recommendation. If the project is not viable, you find out now, not eight weeks in.

Architecture

3–5 days

What we do

Model selection based on your latency, cost, and accuracy requirements
Retrieval strategy design (chunking, embedding, re-ranking)
Tool call and agent loop design
Evaluation harness design, how we know it's working

What you get

A full technical architecture document. Every decision is written down with the trade-offs. Your engineers can read it and understand why we made each choice.

Build

3–8 weeks

What we do

Iterative one-week sprints against the architecture doc
Live working demo every Friday, not a status update
Async Slack updates throughout the week
Eval harness running continuously against each sprint

What you get

Working software every week. By the end of the sprint cycle you have a fully tested, production-ready system with eval scores you can actually read.

QA & Deployment

3–5 days

What we do

Load testing under realistic concurrency
Red-teaming for prompt injection and edge cases
Production deployment to your infrastructure
Monitoring dashboard and alert setup

What you get

A live system with a monitoring dashboard, a runbook, and a handoff meeting where we walk your team through everything.

Maintenance

Ongoing, optional

What we do

Weekly eval runs to catch model drift
Model updates as better options become available
Incident response within 4 business hours
Monthly performance report

What you get

A system that improves over time. Optional, month to month. You can end it any month, no lock-in.

What we commit to.

Fixed price range agreed before we start

We quote a price range after Discovery. We don't move the goalposts mid-project. If scope changes, we discuss it before any extra work starts.

Weekly working demos

Not status updates. Not slide decks. Every Friday you see the software running against real data. If something is off, you tell us Friday and we fix it Monday.

Full documentation and handoff

At close: architecture doc, runbook, deployment guide, and a walkthrough session. Your team can maintain the system without us, and we verify that before we call it done.

Why it works this way.

Questions about how we work.

How long does a typical project take start to finish?

Most AI agent and RAG projects run 6–10 weeks from kickoff to production deployment, including Discovery, Architecture, Build, and QA phases. SaaS MVPs typically run 8–14 weeks. We give a firm timeline estimate at the end of the Discovery phase.

Can we skip the Discovery phase if we already know what we want?

We still run a condensed version. Discovery is where we audit your data, define success metrics, and confirm the architecture will work before we write production code. Skipping it adds risk to both of us. It costs one week and is priced into every engagement.

What if the project scope changes mid-build?

Changes happen. We discuss the impact on timeline and price before doing the extra work, never after. If a change is small, we absorb it. If it shifts the scope materially, we send a one-page addendum with the adjusted range for your approval.

What do you need from us during the project?

Less than most teams expect, but it is specific. We need access to the systems the work touches, a named decision-maker who can answer questions within a day or two, and an hour at the Friday demo. Discovery is the most input-heavy phase because that is where we learn your context. After that, the demos keep you in the loop without pulling your team off their own work.

Who owns the code and the models when the project ends?

You do, completely. Code is delivered into a repository you control, with IP transferred at project close. We build on model accounts and infrastructure in your name, so there are no credentials or licenses held hostage and no per-seat runtime fee to us. If you never speak to us again after handoff, the system keeps running.

Do you sign an NDA and work inside our security requirements?

Yes. We sign NDAs as a matter of course and work within your security and compliance constraints, including BAAs for healthcare data and deployment inside your own cloud tenancy where policy requires it. We raise those requirements during Discovery so the architecture is designed around them rather than retrofitted later.