AI product development and custom SaaS

We build AI products: the agent that does the hard part, and the software around it that makes it usable, safe and yours. An internal tool, a customer-facing product, the thing no off-the-shelf platform quite does. Most builds run two to six weeks, and the code is yours at the end.

The two we point at are both our own.

πCounter is a product we built and run: five agents watch a storefront, its orders, its customer conversations, its ads and its money, write to one record, and bring the owner the change with the evidence attached. The second is a hiring system we built for ourselves, because we needed it.

What makes it an AI product rather than a tool with a chatbot bolted on?

The model does the job the product exists for, not a feature in the corner.

A tool with AI bolted on works fine without the AI and is slightly nicer with it. An AI product has a judgment at its centre: reading a message and working out what someone wants, scoring a candidate against a role, noticing that today's numbers are not like yesterday's and saying why. Take that out and there is no product left.

The rest is the software that makes that judgment safe to rely on. Permissions, an audit trail, the screens a person uses, a boundary on what it may decide alone. That part is unglamorous and it is most of the build, and any studio telling you otherwise has not shipped one.

What have you actually built?

The hiring system is the clearest one, because we ran it on ourselves.

We were hiring a digital marketer. An ad pointed at a landing page where an applicant chose the role and submitted their name, phone number, LinkedIn URL and CV. The system pulled the profile, scored it against the job description, and where the score cleared the bar it phoned the candidate, ran the pre-screen and came back with a transcript and a score.

Nobody paid us to build that as a client project. We needed it, so it exists, and it is the honest version of what we mean when we say the same engine confirms an order, qualifies a lead and screens a candidate.

πCounter is the larger one and the one you can go and look at. It is a separate brand with its own site and its own enquiries, and it is ours.

How long does it take?

Two to six weeks for most builds, and the variable is not the code.

What moves it is how well you can describe the judgment you want made. A decision somebody can talk through end to end without stopping is a fast build. One where three people each describe it differently is a slow one, and the slow part is the conversation that settles it, which happens before anything is written.

We hold the first two weeks after handover under live monitoring, which is how we work.

What happens if you disappear?

You keep the code, the documentation and deployment access, and anyone competent can pick it up.

This is worth asking any small studio and it usually gets answered vaguely, so here it is plainly. The repository is yours. The documentation is written for somebody who has never met us. It runs on infrastructure we manage, and a yearly fee covers maintenance, hosting and the AI usage behind it. Stop paying and you keep everything, it simply stops running until you host it somewhere else.

Nothing in that takes the work back off you, and nothing about it depends on us still being interested.

Where does an agent not belong?

Anywhere the wrong answer is expensive and the right one is a rule.

Pricing, payroll, anything with a regulator attached, anything where a person will be blamed for the outcome. Those get written as logic, tested, and left alone. The fastest way to lose confidence in a product is to let a model make a decision a spreadsheet would have made correctly every time.

Tell us the judgment your business keeps making by hand, and we will tell you whether it should be a product or a rule.