AI features
An AI feature that does one defined job inside your product, passes a test set built from your own records, and has its spend per user capped before launch.
In Melodio, a text message goes in and a finished song with lyrics, melody and vocals comes out: one job, a clear input, a clear output. In a B2B product or internal tool, the job is more often a call summary, a drafted reply or an answer from your own documents. We test models against real examples before choosing one, and design what the user sees when the answer is wrong.
For Founders and product leads adding AI to a web app, and operations leads who want summaries, drafting or search inside an internal tool or CRM.
You need this when
- 01
Your board wants AI in the product, and nobody has said what it should do.
- 02
Your team pastes customer notes into ChatGPT by hand, one record at a time.
- 03
The AI demo looked right on hand-picked examples and gives different answers on real data.
What’s included.
Use-case shortlist
The tasks AI could take on, ranked by time saved and the cost of a wrong answer, with one picked to build first.
Generation
Text, images, audio or music produced from a user’s input, inside your product’s own screens.
Summaries and drafts
Call notes, tickets, threads or records summarised, and replies or reports drafted for a person to check and send.
Search over your data
Questions answered from your documents and records, with the source shown next to each answer.
Evaluation set
A fixed set of real examples the feature is tested against before launch and after every model change.
Guardrails and costs
Limits on input, output and spend per user, a log of every call, and a fallback when the model fails.
How it runs.
A prototype on your own data runs as a Sprint of 1–4 weeks; the production feature follows inside a Build.
- 01 · Pick the job
One task with a clear input and output, chosen with you from the ranked shortlist.
- 02 · Build the test set
Real examples from your data, each with a written description of a good answer.
- 03 · Prototype and compare
Models from OpenAI, Anthropic and Google run against the test set, with cost per run recorded.
- 04 · Build into the product
Interface, guardrails, logging and cost limits, released to staging first.
- 05 · Watch it live
Errors, costs and user corrections reviewed after launch, and the test set re-run on every model change.
Deliverables
- Ranked use-case shortlist
- Prototype on your own data
- Evaluation set with pass criteria
- Production feature inside your app or tool
- Prompts and model settings under version control
- Usage, cost and error logging
- Docs: how the feature works and how to change the model
Works with
What buyers ask.
Will our data be used to train the model?
We use the providers’ paid APIs and check each one’s current data terms in the scope, including training and retention. Fields you mark as sensitive are removed or masked before anything is sent.
What if it gets an answer wrong?
It will, sometimes, so we design for it. Answers show their source, people approve drafts before they go out, and the evaluation set catches regressions when a model changes.
Which model do you use?
Whichever passes the test set at a cost per run you accept. The model sits behind one layer in the code, so changing provider doesn’t mean rebuilding the product.
Pairs well with
Roles, screens, data and integrations written down and ranked before any code.
Let’s build your pipeline.
Tell us what you sell, who buys it and roughly what a customer is worth to you. We reply with questions, then a written proposal.