An assistant in your product
A coach, guide, or support assistant that has to answer from each user’s own data: their results, history, orders, or account.
Test a real AI assistant on your own data.
A fixed-scope pilot for teams that want an assistant, coach, or copilot in their product and need to know whether it will be accurate, safe, and affordable. I build a working assistant on your data and systems, measure it on the questions your users actually ask, and hand you a staging version to test.
- Typical duration
- 2–3 weeks
- Indicative investment
- from €9,500 excl. VAT
- Primary outcome
- A working, evaluated assistant in staging
- Systems covered
- Chat, voice, tools, and retrieval
Why assistant demos stall
An assistant that looks good in a demo often fails on the questions real users ask. It invents a number, answers outside its remit, takes too long to reply, or costs more per conversation than the product can carry. Those problems rarely show up until real data and real questions are involved.
The pilot brings them out early. The assistant reads facts from your systems through tools instead of guessing, every reply is checked before anyone sees it, and an evaluation set built from your own questions shows where it is right, where it is wrong, and what each answer costs.
When a pilot helps
An internal copilot
A tool that helps your team search documents, fill in records, or prepare reports from the systems they already use.
A demo that does not hold up
A prototype that handles easy questions but invents facts, ignores its instructions, or is too slow or expensive to ship.
What the pilot covers
Grounded answers
- Tools that read your data and APIs
- Retrieval over documents and knowledge bases
- Facts and numbers checked against their source
- Conversation memory with clear limits
Safety and control
- Scope, tone, and refusal rules
- Checks on every message and reply
- Fixed safe replies for sensitive topics
- AI disclosure, consent, and age rules where needed
Evidence and running cost
- An evaluation set from your real questions
- Latency and cost per conversation
- A trace of every step in every reply
- Model choice, including EU-hosted open models
What you receive
- A working assistant in a staging environment your team can use
- Connections to the data sources and systems agreed in scoping
- Guardrails and a written policy for what it may and may not answer
- An evaluation report: accuracy, failure cases, latency, and cost
- An optional voice mode with spoken replies
- A recommendation on models and hosting, including EU-only options
- A production plan with running costs and a build estimate
- Code, prompts, and documentation that you keep
How the pilot runs
- 01
Define the job
Agree who the assistant serves, what it may answer, which sources it may use, and the questions that decide success.
- 02
Connect and build
Connect the agreed data through tools or retrieval and build a thin assistant end to end, with guardrails from the first version.
- 03
Measure
Run it against your evaluation questions, fix the failures that matter, and record accuracy, latency, and cost.
- 04
Decide
Hand over the staging assistant and the evidence, with a clear go or no-go and a plan for production.
How I build assistants
Facts come from your systems
The model writes the sentences; the facts come from tools that read your data. A reply that states something no source supports is stopped and replaced before anyone sees it.
Models stay replaceable
Every model call goes through one gateway, so the model can be changed by configuration and re-checked against the same evaluation set. You are not locked into one provider.
Before we start
Which models do you use?
Whatever fits the job, your data rules, and the budget. That can be a commercial API or an open model hosted in the EU. The pilot compares the realistic options on your own evaluation questions, so the choice rests on evidence rather than preference.
Can it work with our existing product?
Yes. The pilot connects to your data and systems through APIs or read-only access agreed in scoping. The staging build has its own chat or voice interface; embedding the assistant in your app is usually part of the production build.
Can the assistant serve young users or sensitive topics?
It can be designed for them, with age and consent checks, a clear statement that users are talking to an AI, fixed safe replies for sensitive subjects, and logs that support a review. We define those rules together in scoping. They complement legal advice rather than replace it.
What happens after the pilot?
You keep the code, the evaluation set, and the report. If the evidence supports it, production work runs as an End-to-End AI Product Build: hardening, scaling, monitoring, and integration into your product.
Other ways to work together
What should your assistant be able to answer?
Send me who it is for, the data it should use, and a few questions it must get right. I’ll reply personally within one business day with an honest view of whether a pilot is the right next step.
Work directly with Jude Mingay, a freelance AI product engineer based in Freiburg, Germany. Remote projects worldwide; English, German, or Portuguese.