M
← All services
02
Package 02
02 / 05

Your users shouldn't have to learn your UI.

I build the assistant that does the job for them — searching, filtering, comparing, acting inside your product.

BOOK AN INTRO CALL →
WHO IT'S FOR

SaaS and marketplace teams with "add AI to the product" on the roadmap and nobody in-house who has shipped it. Best fit when your product has real actions worth automating, not just content worth summarising.

WHAT YOU GET
  • A tool-calling agent wired into your actual product actions — it queries your data, filters, compares, navigates and completes tasks on the user's behalf
  • Streaming chat UI that matches your design system, or headless if you have your own
  • A tool layer over your existing API — typed, tested, with guardrails on what the agent may and may not do
  • An evaluation set and logging, so you can see what it answers and where it fails
  • Deployed to production, with your team walked through how to extend it
HOW IT WORKS
  1. 01

    Discovery call

    What job should the agent actually do, and which product actions matter.

  2. 02

    Scoping doc

    The two to four tools worth building first, plus success criteria.

  3. 03

    Build

    Agent, tools, UI, guardrails, evals.

  4. 04

    Production + handover

    Deploy, monitor, document.

EXAMPLES
  • Ewyn→

    Two agentic loops inside ewyn.ai, built on the Vercel AI SDK with Claude. Brand extraction is a multi-stage pipeline — discover links, classify, pick pages, screenshot, extract colours, fonts, logo and tone into a brand kit every template inherits. Template generation and editing run as a tool-calling loop that acts on the email document through typed tools (replace, insert, remove a component, ask the user a question, finalise), with each call applied to the live draft as it streams. Every send is still triggered by a person or an API call — nothing runs unattended.

  • VenuePlace→

    The venue-discovery assistant on venueplace.com, built with the Vercel AI SDK on Claude Haiku 4.5. A user describes an event in Estonian or English; the agent asks structured follow-ups, searches the live inventory including availability, renders matches on the page, builds a comparison shortlist, answers "does it have X" by quoting venue text, pre-selects rooms, catering and extras, and hands the user a pre-configured booking page. Eight typed tools over the product's real actions, answers grounded only in tool results, rate and token caps — and it deliberately cannot book or pay.

NEXT STEP

A 30-minute call to see whether this is the right shape for your project — no deck, no pitch.

BOOK AN INTRO CALL →
info@mikkoun.com← All services
Mikk Õun — Freelance Product Engineer
info@mikkoun.comPortfolio →