Company
We document ourselves the way we document our models.
Every model we ship comes with a model card: what it is for, what it is not for, how it was tested, and where its limits are. This one is ours. CraftsAI designs, builds and runs AI systems for businesses.
CraftsAI
- Name
- CraftsAI
- Type
- AI development services
- Built for
- Businesses with a workflow worth handing to AI
- Inputs
- Your workflow, your data, and your bar for “good”
- Outputs
- Agents, assistants and models, and the evals that prove them
- Oversight
- People in the loop on every judgement call
- Data
- Stays in your environment
- Licence
- You own the code
Reviewed at the start of every engagement.
§ 1Intended use
Built for teams with real work to hand over.
We are a good fit when at least one of these sounds familiar.
A workflow that eats hours
Your team reads, sorts and replies to the same things every day. We build an agent that takes the routine cases and hands the rest to a person.
Knowledge scattered across documents
Policies, contracts, wikis and old tickets that nobody can search well. We build an assistant that answers from them and shows its sources.
A product that needs an AI feature done properly
You know what the feature should do. We build it into your web or mobile app, with the evals, monitoring and cost controls that keep it good.
Leaders who want to know where AI pays off
Before you commit a budget, a two-week sprint maps where AI helps and what it would cost, and ends with a working prototype.
§ 2Out of scope
What we won’t do.
A good model card is honest about its limits. Ours is too.
Build a demo that can’t reach production.
Won’t doIf your security, data or reliability needs rule it out, we will say so in the first week.
Train on your data without your consent.
Won’t doYour data is used to build your system, in your environment, and for nothing else.
Promise accuracy we haven’t measured.
Won’t doWe quote numbers only after testing on your real cases, and we show you the test.
Replace the judgement calls your people should make.
Won’t doLarge refunds, legal wording, anything about a person’s job: the AI prepares the case, and a person decides.
§ 3Evaluation & principles
How we hold ourselves to account.
Four rules we apply to every system we ship.
Measured, not guessed
Every system ships with evaluations you can read: real cases, a bar we agreed on, and a result you can check.
Your data stays yours
Private deployments inside your environment. No training on your data without your consent.
Model-agnostic
We pick the model that fits the job, and build your system so the model can be swapped when a better one arrives.
People in the loop
The AI handles the volume. Anything that needs judgement waits until a person approves it.
§ 4Two layers
Every task has two layers.
On top, what your team sees. Underneath, what the system actually did. Move the lens over the example, or use the button, to see both.
A customer writes in: the order arrived damaged, and they would like a refund.
The assistant looks up the order and confirms what happened.
It finds the refund policy that applies.
It drafts a reply, and checks the draft before anyone sees it.
The amount is over the team’s limit, so it asks a person to approve.
A person approves. The customer gets the refund and a clear answer.
Tokens
intent = refund_request · damaged_itemTool call
orders.lookup(order_id)Retrieval
refund_policy.md › damaged goodsEval
tone ✓ · policy ✓ · amount > limitHandoff
route → support_lead + summaryLog
approved · reply sent · case → eval set
People set the goals and make the calls. The machine does the volume, and shows its working.
§ 5Maintainers
Help us maintain this card.
We hire engineers and designers who care as much about measuring a system as building it.
- Maintainers
- A small team of engineers and designers, and maybe you.
Put us to work on one workflow.
Pick the workflow that costs your team the most time. We will come back with an honest view of what AI can do there, how we would measure it, and what it would cost.