One workflow. One agent. Running in three weeks.

You pick the repetitive job that eats your team's time. I build a Claude-based agent that does it, prove it works with evaluations, deploy it where you want it, and hand it over with documentation. No retainer, no discovery phase that never ends.

Scope your sprint ↓ Email me directly

Scope §1

What should the agent take off your team's plate?

Pick the closest match. The statement of work fills in with your choice, and the button sends it to me as an email you can edit before it goes.

Deliverables §2

What you have at the end

A working agent in production

Deployed in your cloud, on your VPS, or on infrastructure I set up for you. Your API keys, your data, your account.

An evaluation suite

A labelled test set built from your real cases and a script that scores the agent against it. You rerun it whenever you change a prompt or a model.

Guardrails and a human path

Confidence thresholds, escalation to a person, logging of every decision, and limits on what the agent is allowed to touch.

Handover you can act on

Source code in your repository, a runbook, cost estimates per month, and a recorded walkthrough for whoever maintains it.

Thirty days of fixes

Anything that breaks or misbehaves inside the agreed scope in the first month, I fix at no charge.

A straight answer on fit

If week one shows the job is not a good fit for an agent, I tell you, and you stop paying. See the exit clause below.

Timeline §3

Three weeks, in order

Week 1Spike

One 60 minute call to pin inputs, outputs and done
Access to a sample of real cases
A rough agent against 20 to 50 of them
Go or no-go with numbers

Week 2Build

Integrations with your systems
Eval suite grows to a few hundred cases
Guardrails, escalation, logging
Mid-week demo on your data

Week 3Ship

Deploy to your environment
Shadow mode next to the current process
Runbook, walkthrough, cost sheet
Handover and start of the 30 day window

Week-one exit. If the spike shows the agent cannot reach the accuracy bar we agreed, we stop. You pay €1,500 for the week and keep the spike code and the findings. No further invoice.

Price §4

Fixed, quoted after a 20 minute call

€4,500
to €7,500per sprint, excluding VAT, invoiced from a Greek sole proprietorship

50% on signing, 50% on handover. Bank transfer or Stripe.
The band depends on the number of systems the agent touches and how messy the inputs are. Most sprints land in the middle.
Not included: model API costs (usually tens of euros a month at SME volume), your cloud bill, and new work outside the scoped workflow. That becomes a second sprint.
Optional after handover: a monthly care plan for monitoring, model updates and small changes, from €600 a month. Never required.

Fit §5

Who this is for

Good fit

Teams of 5 to 200 where one person spends hours a week on the same text-heavy task
You can give me real examples of the job, including the awkward ones
Someone on your side can answer questions within a day during the sprint
You want the code and the keys, not a subscription to a black box

Not a fit

"Add AI to the product" with no specific job in mind
Decisions with legal or medical consequences and no human in the loop
Jobs where the inputs change every week and nobody can say what correct looks like
Chatbots for the sake of a chatbot

Who builds it §6

Harry Kaonis, Athens

Senior software engineer, seven plus years shipping production software across web, mobile, backend and infrastructure. I work alone on each sprint, so the person you talk to on the call is the person writing the code.

Built agent pipelines with evaluation harnesses, blinding and regression checks for an AI product team
Built React and Next.js web apps and internal dashboards, from requirements to production release
Shipped consumer and enterprise React Native apps to the App Store and Google Play, with CI and crash monitoring
Run production infrastructure: Docker, reverse proxies, TLS, backups, monitoring, Cloudflare, Hetzner
Launched goev.gr, a 600 page site with an API, daily data pipelines and automated content, launched in a single night and extended daily

Questions §7

The things people ask first

Which model, and can we use something other than Claude?

Default is Claude through the Anthropic API, because the tool-use and long-context behaviour is the most predictable for this kind of work. If you need an EU-hosted or open-weights model for policy reasons, say so on the call. The evals tell us if it holds up.

Where does our data go?

To your model provider account under your own data processing terms, and to infrastructure you control. I work on a sample during the sprint and delete it at handover. A short data processing agreement is included in the contract.

What if we have no API for the system the agent needs to touch?

Most SaaS tools have one. For the ones that do not, we fall back to email, CSV exports, or a shared inbox the agent reads. That is usually fine for a first version and we discuss it in week one.

Can you do this for a web app, a mobile app or an internal tool instead?

Yes. The same fixed-scope format applies to React and React Native work and infrastructure setups. Email me what you have and I will tell you whether it is a sprint or a smaller job.

How many sprints can you run at once?

Two. I keep capacity deliberately small so each sprint gets daily attention. Start dates are first come, first served, and a signed deposit holds your slot.