AI that does real work in your business
A model tuned to the job you need done, AI features inside the product you already have, or workflows that run on their own. Tested on real examples, built with guardrails, at one fixed price. And first, a straight answer on whether AI is the right tool at all.
Three kinds of AI work
Each starts with your use case and a test set, so you see it working on your own examples before it goes anywhere near a customer. Pick the one that sounds like your problem, or read all three.
- Custom and fine-tuned modelsSorting tickets, emails or documents · Pulling fields out of invoices and formsSee the workflow
- AI features in your productAssistants that answer from your data · Search that understands the questionSee the workflow
- AI workflows and automationsInbox triage and drafted replies · Invoices and receipts read and filedSee the workflow
Custom and fine-tuned models
A model trained or tuned on your own examples to do one narrow job very well: sort, extract, score or write in your house style. We only tune when a well-instructed off-the-shelf model falls short, and we prove the difference on examples it never saw.
- Sorting tickets, emails or documents
- Pulling fields out of invoices and forms
- Scoring leads, risk or quality
- Writing in a set tone or format
How it runs
The job, in one sentence
What the model decides, from what input, and who uses the answer. If we can’t write it down plainly, it isn’t ready to train.
Your examples, gathered and checked
Real cases of the job done well, cleaned and labelled with you. A slice is set aside as a test set the model never trains on.
A baseline first
An off-the-shelf model with good instructions, scored on the test set. Often that’s enough, and you’ve saved the cost of tuning.
Tune or train, only if it earns it
When the baseline falls short, we tune a model on your examples and score every version against the same test set.
You see the numbers, then decide
A short written result: how often it’s right, where it fails, and what it costs per thousand runs. You choose what ships.
Running in your accounts
Hosted where you choose, logged and monitored, with a plan for retraining as your data changes.
What you get
- A tuned or custom model, or proof you don’t need one
- The test set and the written scores
- Hosting in your accounts, with monitoring
- A retraining runbook
What we need from you
- Real examples of the job done well (we’ll say on the call how many it needs)
- Someone who can judge whether an answer is right
AI features in your product
Intelligence inside the product your customers already use: an assistant that answers from your own data, search that understands the question, documents read and summarised, replies drafted for a person to send. Designed into your screens, not bolted on beside them.
- Assistants that answer from your data
- Search that understands the question
- Summaries of long records and threads
- Drafts a person reviews and sends
How it runs
Find the moment
Where in your product do people waste time or get stuck? We watch the flow and pick the one place AI helps most.
Design it into your screens
A clickable design that fits how your product already works, approved by you before any model is called.
Ground it in your data
Answers come from your own records and documents, with the source shown, so people can check them.
Set the limits
What it may do alone, what needs a person, what it says when it doesn’t know, and a spending cap that holds.
Ship to a few users first
Behind a switch, to a small group, with every request and answer logged so we can see how it really behaves.
Roll out and keep watching
Wider release once it holds up, with quality and cost tracked over time and an alert if either drifts.
What you get
- The feature, live in your product
- Guardrails, spending caps and fallbacks
- Logs and a quality dashboard
- Documentation for your team
What we need from you
- Access to the codebase, or the developer who owns it
- The data the feature should answer from
AI workflows and automations
The repetitive work behind your business, done for you: inboxes read and sorted, invoices filed, forms filled, follow-ups drafted, reports written. Connected to the tools you already pay for, with a person approving wherever a mistake would cost you.
- Inbox triage and drafted replies
- Invoices and receipts read and filed
- Lead follow-up and CRM updates
- Weekly reports written for you
How it runs
Map the work as it is today
One walkthrough with the person who does it now. We write down every step, every tool and every judgement call.
Pick what AI should do
Some steps suit AI, some suit plain automation, and some should stay with a person. We mark which is which.
Connect your tools
Email, CRM, spreadsheets, accounting, chat. The workflow works inside the systems you already use.
Put a person where it matters
Approval points before anything costly or customer-facing goes out, with one click to approve or fix.
Log it, alert on failure
Every run recorded. When something fails or looks unusual, a person hears about it before a customer does.
Hand over with a runbook
A plain guide to what it does, how to pause it, and what to do when it asks for help.
What you get
- The workflow, running on your accounts
- Approval points where they matter
- Run logs and failure alerts
- A runbook your team can follow
What we need from you
- Access to the tools involved
- An hour with whoever does the work today
What makes it safe to rely on
A demo that answers one question well is easy. A feature that still behaves on the ten-thousandth input, at a cost you can predict, is the real job.
A straight first answer
Some problems don’t need AI, and some need a model that doesn’t exist yet. We say so on the call, before you spend anything.
A test set before a model
We agree in writing what “working” means and score every version against real examples, so the result is a number, not a feeling.
Guardrails and fallbacks
What the AI may do, what it must hand to a person, and what happens when the provider is down or the answer is uncertain.
Costs you can predict
Usage caps, caching and the right model size for the job, so a feature that shines in a demo doesn’t surprise you in month two.
Your data stays yours
Provider settings that keep your data out of training, keys in your own accounts, and a written note on what goes where.
Logged and monitored
Every request and answer recorded, quality tracked over time, and an alert when it starts to drift.
From use case to production
Start with the use case alone, and decide once you’ve seen the prototype work on your own data.
Use case
What the AI should do, what data it can learn from, and how we’ll measure whether it works.
You getA scope, a fixed price, and a straight answer on whether AI fits
Prototype
The smallest version that proves the idea, scored against real examples before anything is built around it.
You getA working prototype, tested on your data
Build
Test sets, fallbacks, spending caps and logging, so it keeps behaving when real inputs arrive.
You getThe model, feature or workflow, with guardrails
Launch
Deployed to your accounts, monitored, and documented for whoever runs it next.
You getRunning inside your product or systems
Support
We stay close after launch, with an optional plan to keep improving it as your data changes.
You getDefects fixed at our cost for 30 days
AI integration questions, answered
Do I need my own model, or will an existing one do?
Usually an existing model with the right instructions, examples and guardrails does the job, so we start there. A tuned or custom model earns its place when the task is narrow, the volume is high, or the data is sensitive. We test both before we recommend one.
Which AI providers do you work with?
The major model providers, and open models you can host yourself. The choice depends on the task, your budget and where your data is allowed to go, and it goes into the scope with the reasons.
Can you add AI to software another team built?
Yes. We read the code first, then design the feature to fit how the product already works. If it’s fragile underneath, we’ll tell you, and that may mean a Ship It job first.
How do you stop it making things up?
We ground its answers in your own data, limit what it’s allowed to do, score it against a test set, and hand uncertain cases to a person. No system is perfect, so we design for the failure, not just the demo.
What does it cost to run afterwards?
The build is one fixed price. Running costs are the provider’s usage fees plus your hosting. We estimate them before you agree to the scope, and cap them in the build so they can’t run away.
What happens after launch?
If something we delivered breaks within 30 days of acceptance, we fix it at our cost. After that, keep us on to improve it as your data changes, or run it yourself with the documentation we leave.
Tell us what you want AI to do
Describe the job in a sentence or two. On the call you’ll get a straight answer on whether AI fits, and what we’d build if it does.