SERVICE · 02
Custom AI development.
You have one specific workflow an LLM should solve. I scope it, build it, and ship working software to your repository in two weeks. Not a demo. A narrow system with an acceptance test.
- price
- €6,000 fixed
- build time
- Two weeks
- output
- live in your repo
- handoff
- + 14-day bug-fix window
- cap
- 1 build at a time
WHAT YOU GET
Six things. Every build.
A scoped system spec
Day one. A document that names exactly what we are building, what we are not, and the eval that proves it works.
Working software in your repo
On your stack — Postgres, TypeScript, Python, your existing infra. Reviewable PRs, not zip files. No Lambda graveyard.
Eval suite and monitoring
Every prompt has an eval. Every eval has an alert. You catch model drift before your users do.
Versioned prompts and ops docs
Prompt history under source control, a runbook for the first 90 days, and a one-pager your team can extend from on day 31.
14-day bug-fix window
I stay available after ship for defects against the agreed acceptance test. New scope becomes a separately approved sprint.
A second opinion before we start
If the plan is wrong, I will say so on the call. I do not do month-two pivots.
DELIVERY
One workflow. Four gates.
Prove AI belongs
Define the user, decision, data, and failure cost. A deterministic rule wins when it is the better tool.
Write the acceptance test
Agree on the eval set, latency, cost ceiling, security boundary, and the human fallback before code.
Ship vertical slices
Reviewable changes land in your repository. Each slice is measured against the agreed eval, not a polished demo.
Run the acceptance test
The final handoff leaves your team with source, prompts, evals, and a runbook. The 14-day window covers defects against the agreed test.
Common builds include RAG systems, internal agents, and eval-first AI workflows. The build-versus-buy decision comes first; read the framework I use before adding another platform.
FIXED SCOPE · FIXED PRICE
€6,000. Two weeks. No expanding estimate.
If the problem does not fit this box, I phase it or tell you the sprint is the wrong delivery model before work starts.
INCLUDED
- ✓One workflow and one primary user role
- ✓One existing data source or API integration
- ✓A written system spec and acceptance test
- ✓Working code delivered to your repository
- ✓Eval set, runbook, and handoff session
- ✓14-day bug-fix window after delivery
SEPARATE PHASE
- —Multi-tenant product architecture or complex auth
- —Data-platform migrations or model fine-tuning
- —Unlimited integrations, support, or changing scope
FAQ
Direct answers.
What can't you build? +
Anything that needs a model I cannot access. Anything that requires fine-tuning on more than 100k examples — I will refer you to someone who specialises in that. And anything where the right answer is "just write a regex".
Why two weeks? Most agencies say three months. +
The sprint is intentionally narrow: one workflow, one existing data source or API, a defined acceptance test, and a production handoff. A larger system is divided into separately approved phases instead of stretching the sprint.
What stack do you use? +
Your stack. I build in TypeScript or Python on Postgres, pgvector, and your existing infra. No new SaaS to subscribe to. No Lambda graveyard to inherit.
What happens after the 14-day bug-fix window? +
Three options: clean handoff, an Advisory or Operate package for ongoing decisions, or another tightly scoped sprint. The default is a clean handoff; there is no required support contract.
New to this? Start with how to hire an AI consultant — what they cost, what they should ship, and how to avoid the wrong one.
Bring the workflow. Leave with the smallest build that can prove it.
A focused 20-minute call to identify the constraint, choose the smallest useful scope, and decide whether I am the right person to ship it.
- 01Diagnose the constraint
- 02Choose the useful scope
- 03Leave with a next move
No pitch deck · no sales handoff
Book a 20-minute intro See the fixed sprint