Jorge Polo

Building AI products · Madrid

Design is a decision, not a deliverable.

Customer, product and AI in one seat. I design AI products end to end, from concept and feature architecture to multi-agent pipelines and prompts in production.

_now

Most companies put three people between a customer and an AI feature: one who talks to the customer, one who defines the product, one who makes the AI behave. At Feedier, a European Customer Intelligence Platform, I’ve held all three.

I joined as a Product Designer in February 2025, as the platform was becoming AI-first. Within a year the role covered product concepts from idea to shipped feature, AI capabilities down to the prompts that run in production and the method that says whether they work, and the enterprise accounts that depend on them.

I also build the tools around the work. To run those accounts I set up my own operating system with Claude and MCP connectors across Slack, Jira, Gmail, Notion and HubSpot: scattered context in, a structured brief out, in one pass.

I work with problems: finding the one worth solving, then solving it. The deliverable is never the point; the decision behind it is.

_work

  • From tree to map

    Feedier · Text analysis · 2025–2026

    Feedier’s analysis engine turns thousands of customer verbatims into decisions. With the Product Lead, I co-authored the model behind its text analysis: flat topics for the AI to tag, flexible segments for people to read. It shipped in July 2025 with an automatic migration for every account. The model held; the taxonomies inside it didn’t. So I designed how an LLM generates taxonomies people can actually read, and a method to diagnose whether a live taxonomy’s tagging can be trusted.

  • Analyst mode

    Feedier 4.0 · UX/UI · 2025–2026

    After years of adding features, the interface reflected the order in which things had been built, not what each user does. I designed three changes: a navigation organised around two journeys, analyst and admin; one analysis view that opens from any object, whether topic, segment, insight, source or attribute; and tracking on action plans, so a team can see whether a problem went down after they acted. Concept to QA, across five releases.

  • MCP before it was obvious

    Feedier · Platform · Ask Feedier

    With the Product Lead, I pushed to expose Feedier’s analysis through MCP, an open protocol, so it could live in Teams, Copilot and internal assistants instead of only inside our UI. The first technical answer was that it wasn’t feasible; Ask Feedier shipped on top of it. Owning the behaviour meant owning the engine: when sparse verbatims failed to surface under broad filters, I traced it to a tool suppressed in the system prompt.

  • Internal tools that give hours back

    AI consultancy · Co-founder · 2026

    Small firms lose hours to manual, repetitive work. We build the system on how they already work, with Claude. For an accounting firm: a tool that turns each client’s accounting export into the firm’s bank-format balance sheet, keeps its subtotal formulas intact and checks every line against what the team would have typed by hand; when an account has no place in the template, it stops and asks. A balance now takes minutes instead of hours. The same discipline runs our eight-agent diagnostic pipeline: an audit layer that checks the system’s own output before anyone relies on it.

Full case studies, with screens, on request.

_lab

Things I build to learn how AI behaves, mostly with Claude Code.

  • Playtest agents · A multi-agent system that simulates full board-game matches to test balance: a game master, one agent per player with its own private hand, and a reporter. Same setup, two rule versions, to see where the balance shifts.
  • Doodle heads · Procedural characters in my own illustration style, drawn entirely in code, with zero image assets.
  • Camino de letras · A handwriting app for iPad and Apple Pencil, so my son can practise joined-up letters: sessions of about eight minutes, no punishments, no engagement tricks.
  • Harmonices · A contemplative strategy game: you compose the solar system, and the music is its physics.
  • History generator · An alternate-history game from 1945: a simulation engine and a chain of prompts build a plausible 2025, with explicit causal memory and AI costs kept in check.
_principles

How I design AI products

  1. Separate what the model does from how a person reads it.
  2. Coverage is not quality: an output nobody can read has failed.
  3. A person approves every write. The system proposes; it doesn’t decide.
  4. Never trust a single metric. Evaluate on the same snapshot, before and after.
  5. What already works is part of the spec. A change ships only if the existing use cases still hold.
_background
  • 2025 – nowFeedier · Product Designer, then AI features, product and enterprise accounts
  • 2026 – nowAI consultancy · Co-founder · AI products and internal tools for small firms
  • 2019 – 2024Idoven, Rive, Kodit, Holalucas · Product and brand designer in health-tech and prop-tech startups
  • 2017 – 2018Boxrun · Co-founder of an urban courier app
  • 2009 – 2020Agencies and freelance · Brand, editorial, web, motion and 3D

Spanish, English, French.

_contact

Building an AI product or tool? Let’s talk.

Message me on LinkedIn

CV and case studies on request.