Service 04
AI and agents
LLM features, assistants on your own data, and agents that take real actions.
Assistants that answer from your documents instead of guessing, and agents that call your APIs to get work done: file a ticket, update an order, draft a report. Built with evaluations and guardrails, so it behaves in production, not just in the demo.
What you get
Assistants on your data
Retrieval over your documents, tickets or database, with citations so people can check every answer.
Tool calling agents
Agents that use your internal APIs to complete tasks, with clear limits on what they are allowed to do.
Evaluations
Test sets and automated checks that measure quality before and after every change to prompts or models.
Chat interfaces
Streaming chat UIs that feel instant, with handoff to a human when the agent should not decide alone.
A good fit for
- +Support teams answering the same questions
- +Companies with large internal knowledge bases
- +Products adding AI features for their users
Typical stack AI SDK, TypeScript, Python, pgvector, LangGraph
How we work
011 week
Discovery
A call and a short workshop to understand the problem, the users and the constraints. You get a written scope, a timeline and a fixed price.
021 to 3 weeks
Design
Flows and screens for the parts that matter most, reviewed with you before any production code is written.
034 to 10 weeks
Build
Weekly releases to a private preview link, so you see real progress and can change direction early.
04ongoing
Launch and support
Launch, monitoring and a support period, then an optional retainer for new features and maintenance.
Related work
Questions
Which models do you use?+
Whatever fits the task, cost and data rules. I design the system so the model can be swapped without a rewrite.
Is our data safe?+
Data stays in your infrastructure where possible, providers are chosen for their data policies, and personal data is masked before it reaches a model.
How do you stop the agent from making things up?+
Answers are grounded in retrieved sources with citations, risky actions need confirmation, and every release is checked against an evaluation set.