Engineering and technical consulting
AI development
Build AI-powered product features with clear boundaries, useful evaluation, and a path from prototype to operation.
Who this is for: Founders and product teams adding an assistant, language-model feature, or tool-connected AI workflow to an existing product.
An AI demo can look convincing before it encounters unfamiliar inputs, restricted data, or a failing integration. I help turn the intended user outcome into an implementation plan: what the model should do, what the surrounding software must enforce, and how the team will decide whether the feature is useful enough to release.
Discuss your projectProblems I can help solve
- An assistant needs access to business tools without unrestricted authority.
- A prototype produces plausible answers, but quality is difficult to assess or reproduce.
- A model integration needs to fit existing authentication, data access, and operational requirements.
Decisions that shape the implementation
Model responsibility
Choose the smallest useful model task and define its inputs and outputs. Keep permission checks, irreversible actions, and business rules in application code; use explicit approval where the consequences justify it.
Evaluation before expansion
Agree on representative examples, unacceptable failures, and acceptance criteria. Compare model and prompt changes against that set, including missing information and tool failures, before broadening the feature.
Cost and operational behavior
Design around latency and usage budgets, timeouts, bounded retries, and a useful fallback. Decide what can be logged for diagnosis without exposing sensitive content.
What an engagement can deliver
- AI feature architecture and integration plan
- Model and tool integrations with structured outputs
- Evaluation examples and acceptance criteria
- Human-review, fallback, and operational handoff flows
How we begin
- Clarify the user task, available data, permitted actions, and success criteria.
- Build a narrow vertical slice with representative inputs and a baseline evaluation.
- Refine quality, failure handling, and operating costs; document the release decision and handoff.
Related engineering work
Kelft: AI-native language design
Explore the requirements, verification contracts, and execution boundaries behind my language and tooling work.
Common questions
Is AI development the same as workflow automation?
No. AI development introduces model-driven behavior such as interpreting language or generating a response. Workflow automation connects deterministic steps and may need no model. I separate those responsibilities so each can be tested and operated appropriately.
Can you work with an existing product?
Yes. An engagement can focus on the model integration, an assistant feature, or the architecture and evaluation of a prototype. The starting point is your existing product, constraints, and intended users.
Bring the problem and the constraints.
We can define a focused build, an architecture review, or ongoing technical collaboration.
Start a conversationHiring for a senior or lead role? Read my engineering profile.