AI Proof-of-Concept Sprint

Two weeks on your own data. You get a working prototype, the numbers it scored against targets you agreed, and a written answer: should this go to production?

Length2 weeks, 10 working days
Built onYour data, under NDA
You keepCode, eval set, report
Price$2,000, fixed

Book a technical consult See the 10 days

README.md

contracts-qa

make setup
make eval    # 260 questions
make serve

Handover notes, architecture, how to run the evals.

Eval set

260 questions, agreed with your team

What is the notice period on the Lyon lease?	Clause 14.2
Who can approve a refund over 5,000?	Annex B
Which parties sign contract 88?	Page 1
Is travel time billable?	Refuse

Go/no-go report · Day 9

Recommendation: go, after two fixes

Measure	Target	Result
Answers correct and cited	90%	94.0%
Unsupported claims	under 2%	1.6%
Cost per answer	under $0.015	$0.009
p95 latency	2.5 s	2.8 s

Fix first: latency, then refusals on scanned annexes.

Signed off with your team, day 10

Ten working days

The eval set comes before the prototype, so every change in week 2 is measured against the same bar.

Days 1 to 2Access and questionsRead access to a sample of your data; the questions it must answer.

Days 3 to 4The eval setReal questions, accepted answers, and targets for accuracy, cost and latency.

Day 5First buildRetrieval and prompts, measured on the full eval set.

Days 6 to 8Build and measureGuardrails, refusals, cost and latency, with the eval set run after every change.

Day 9The go/no-go reportWhat passed, what is open, what production costs.

Day 10The reviewGo, go after fixes, or no-go.

What you keep after day 10

All three are yours, whatever the answer is.

The prototype

Working code on your data, in your repository, with the eval harness beside it.

Prototype mergedday 10In your repository, on your main branch
Eval harness addedday 10make eval runs all 260 questions again
Sandbox access handed backday 10Our keys removed from your accounts

The eval set

Every question, its accepted answer and every run, so the next team measures against the same bar.

Answer accuracyLast 14 eval runs

94.0%6 ptssince run 1

The go/no-go report

A written recommendation with the numbers behind it, and the fixes production needs first.

Go/no-go report, eval run 14

Recommendation: go, after two fixes

The prototype answered 94.0% of 260 test questions correctly, against a 90% target.

Cost per answer is $0.009, under the $0.015 target. p95 latency is over target and is fixed first.

Refusal when unsure is at 84% against 90%. The build starts with both fixes and the same eval set.

Next step: Product Build, 8 to 16 weeks.

Three possible answers

A no-go after two weeks costs far less than a failed launch after six months.

Go

The prototype met its targets. The next step is a Product Build of 8 to 16 weeks, using the same eval set.

How a build runs

Go, after fixes

Most targets met, a few still open. The report lists the fixes, and the build starts with them.

See the build prices

No-go

The numbers say not yet. You keep everything, and the report says why, so you can decide what to change.

How we measure

What we need from you

Four things, agreed on the consult before the sprint starts.

A sample of real dataDocuments, tickets or records
One domain expertTo agree the eval set
AccessA sandbox, or we host it
The decision makerAt the day 10 review

Is our data safe during the sprint?

We sign your NDA before we see anything, work in a sandbox you control or one we set up for you, and delete our copies at handover.

Which models do you use?

Whichever scores best on your eval set for its cost: OpenAI, Anthropic or an open model you host yourself.

What if our data is messy?

That is normal, and the sprint will show it. The report says what the data needs before production.

Can the sprint run on our own infrastructure?

Yes. We can build inside your cloud account so nothing leaves it.

What does it cost?

$2,000, fixed, agreed before day 1. Every price is on the pricing page.