Carl Palmquist · Stockholm

I design and build AI products, end to end.

Renofine, Produlog and Concept Forge are live below. Before that, four years at Epidemic Sound on the Partner API. Stockholm, open to the right team.

Selected work

2026 · Founder · Product · Engineering · renofine.com ↗

Renofine

Renofine on an example flat: the floor plan with tabs for rooms, budget, purchases and the client view Try the app live
Renaida, the AI colleague: describe a job, get a project with rooms, tasks and a budget Try Renaida live

The example flat in the real app. Drag a wall and watch the numbers change, or open Budget and 3D. Open the example flat ↗

Software for planning and running a home renovation. Describe the job or draw the flat, and Renaida, the AI colleague, turns it into rooms, tasks, a budget and the Swedish tax deduction. The builder, the homeowner and the tradesperson each get their own screens. Ten languages.

What I had to get right

  • Evals for every AI feature. Each one runs against a test set I labelled by hand. A receipt photographed sideways once came back confidently wrong: a different shop, a different total, at 0.72 confidence, and nothing on screen showed it. The evals catch that kind of error.

  • The AI suggests, the user approves. Drop a quote PDF in, get a project draft. The AI proposes the changes, you approve them side by side with what is there today, and every fact links back to the document it came from.

  • Locked-down database access. Two simple checks decide what each user can see, and the AI reads data as the signed-in user, never with admin keys. Every database change gets tested from two angles: the owner's view and an invited user's view.

  • Three users, three sets of screens. Each role gets its own screens, so a change for one role leaves the others alone.

  • Floor plans in real-world units. The drawing tool works in millimeters, not pixels. Zoom in, zoom out, change a setting, and the math still works because the units are real.

  • Ten languages kept in sync. English is the source. Swedish is required. The others are best-effort. No text is hardcoded into the screens.

Built with: React and TypeScript for the screens. Supabase for the database, login, and file storage, with 370+ versioned database migrations. Konva for the drawing tool. Cloudflare to host it. Claude and OpenAI for the AI helpers. A type checker that is not allowed to get worse.

2026 · Founder · Product · Engineering · getprodulog.com ↗

Produlog

Kaisa, the catalog assistant: give it a shop address, a vibe or a messy spreadsheet Try Kaisa live
Three demo brands in Produlog: Maja Linen Studio, Nordbrenn Lighting and Studio Argile

Kaisa builds a catalog from a shop address, a vibe or a messy spreadsheet.

Paste a shop address or a spreadsheet and Produlog builds the catalog: products, prices and images in the brand's own fonts and colours, out as a PDF, a live link or a wholesale linesheet. Built for fashion and home brands.

What I had to get right

  • AI drafts the product descriptions, and the brand edits them in its own voice.

  • Every output is editable in one click, so a wrong AI draft is quick to correct.

  • One source, many outputs. The same product data comes out as a PDF catalog, a live link or a wholesale linesheet.

Built with: React and TypeScript for the screens. Supabase for the database, login, and file storage. Cloudflare to host it. Claude and OpenAI for the AI helpers.

2022 to 2025 · Partner API Manager, then Customer Success Manager

Epidemic Sound, Partner API

Music Catalog API v2

Tracks

GET /tracks/{track_id}

Parameters

🔒 bearer

Name Value
track_id *
string · path
include
array · query
GET /tracks/search
GET /tracks/{track_id}/stems

Licensing

POST /licenses
DELETE /licenses/{license_id}

An illustration of the kind of API I packaged and demoed, with an invented catalog. Not Epidemic Sound's docs. Pick a track and hit Execute.

The Partner API powers Epidemic Sound's integrations across creator tools and editing platforms. I helped shape it and championed it internally, on the commercial and partner-facing side: packaging, partner deals, demos and onboarding, alongside the product and engineering team that built it. It grew past ten million monthly users in those years.

What I did

  • Commercial packaging. Translated technical capability into pricing tiers, contract terms, and value propositions partners could understand and buy.

  • Live demos. Walkthroughs with platform leads at editing tools, content platforms, and creator software. Knowing the product end-to-end meant I could answer technical questions on the spot, without escalation.

  • The partner voice in the roadmap. Brought partners' real needs, the engineering cost and the risks to the roadmap and pricing decisions, and argued for saying no to a lot of requests so the API stayed sharp.

  • Customer success at scale. Onboarded partners from cold inbound to multi-million-user integrations. The companies that scaled with us had a working partnership behind the integration.

  • The bridge between commercial and engineering. Translated partner pain into engineering work, and engineering trade-offs back into language commercial leaders understood.

Cases

Concept Forge Try it live

Slot games · Creative AI pipeline

Concept Forge

Type a vibe and get a complete slot game concept: name, key art, a playable reel, a soundtrack and sound effects. Open Concept Forge ↗

Partner Feed Health Try it live

Micromobility · Partner data · Built in a week · Public data only

Partner Feed Health

Voi's public data feeds, checked live against what Google Maps, MaaS apps and cities need. Open Feed Health ↗

Built for myself · Product · Engineering · Used daily

My own operating system

The clearest proof of how I'd build a self-improving loop for a team is the one I run on myself. All my tasks, notes and documents live as plain text an AI can read and act on, with a small app sitting on top of it.

What I had to get right

  • Everything the system needs lives as plain markdown in one place, including the corrections I've given it over time.

  • When I drop in a screenshot of something to handle, it gets read, drafted into the right document and filed where it belongs, so the analysis and the legwork happen without me.

  • The app is disposable and I rebuild it in an afternoon when my needs change, but the markdown underneath stays the same, because that's where the value is.

  • It prepares and drafts everything, but I make the last call on anything that gets sent, paid or published.

Built with: Node with a thin front end I rebuild on demand, plain markdown for all the data, Claude Code as the engine, and a few APIs like Calendar wired straight in. It runs entirely on my own machine.

About

Carl Palmquist

Hi, I'm Carl.

Four years at Epidemic Sound across partnerships, product and customer success. In the Partner API team I ran partner onboarding, the feedback loop into engineering, and helped shape the commercial models that priced the platform. It grew past ten million monthly active users during that stretch.

These days I run my own products, Renofine and Produlog, and design and build every part of them myself. I build, I package, I demo, and I sit with the people who end up using it.

I don't write most of the code by hand. I decide what gets built, write the checks that show whether it works, and own what goes live.

  • Stockholm
  • Swedish and English
  • Open to the right team

How I build

  • 01

    Plans that stay alive

    Every project has one file with this week's plan, what shipped, what's stuck, and open questions. The AI reads it before every session.

  • 02

    Save often. Ship deliberately

    Every change saves to a local copy automatically. Every database change gets verified automatically. Putting it live is my decision.

  • 03

    Build the check into the feature

    AI made building five to ten times faster. The bottleneck moved to knowing what's worth building and whether it worked. So every feature ships with its own check.

Three more ↓
  • 04

    Save what you learn

    When the user says 'don't do that', I write it down. When a design choice turns out wrong, I write that down too. The next session reads the notes and skips the same mistakes. The expensive part of working with AI is re-explaining the context every morning, so I write it down once.

  • 05

    Rules in plain text

    A short file at the top of every project. Things like: no shortcuts on types. No file longer than 500 lines. Every database change comes with an undo script before it runs. The AI follows the rules when they are written down.

  • 06

    Five advisors in text files

    A CEO, a designer, a tech lead, a small business owner, a homeowner: each one is a small text file describing how they think. Before a big decision the AI asks each one. They argue, I read it and pick. It costs nothing and catches blind spots.

Get in touch

Stockholm. Open to the right team.
Say hi.