Charlie Holtz on Rapid AI Prototyping

Table of content

Charlie Holtz builds AI demos that go viral. His “Narrator” project—which uses GPT-4 Vision and ElevenLabs to have David Attenborough narrate your life through your webcam—made headlines when Business Insider reported it “struck fear into the heart of Hollywood” over concerns about AI voice cloning.

The project took a weekend to build. It has 4,400 stars on GitHub.

From Quant Finance to AI Hacker

Holtz worked in quantitative finance at Point72 before joining Replicate as their first Hacker in Residence. The role fit his style: build interesting AI demos quickly, ship them publicly, learn what resonates.

At Replicate, he created:

CHARL-E: The first Mac app for running Stable Diffusion locally on M1 chips
StickerBaker: AI sticker generator built in Elixir
EmojiGen: Custom emoji creation that hit #1 on Hacker News
Zoo: A text-to-image playground for comparing models

Each project shares a pattern: take a new AI capability, wrap it in a simple interface, ship it fast.

The Build Fast, Ship Public Philosophy

Holtz’s workflow is aggressive iteration. Rather than planning for weeks, he builds something in days, posts it on Twitter/X, and watches what happens. His highlights feed reads like a demo reel of experiments.

The approach works because:

Quick feedback loops - Ship early, see what resonates
Composable AI primitives - Combine vision models, language models, and TTS into novel applications
Simple interfaces - Most demos are one function, one button, one result

From Demos to Products: Melty

In 2024, Holtz co-founded Melty Labs with Jackson de Campos (previously at Netflix building ML infrastructure). Melty is a “chat-first code editor” with a specific twist: every chat message is a git commit.

The idea emerged from frustrations with existing AI coding tools:

“We’ve used most of the AI coding tools out there, and often ended up copy–pasting code, juggling ten chats for the same task, or committing buggy code that came back to bite us later.”

Melty integrates AI into the version control loop itself. You can revert, branch, reset, and squash your AI conversations just like commits. The AI watches your changes, so you don’t have to explain context—it already knows.

Within 28 days of building, Melty was writing about half its own code.

Current Work: Conductor

Holtz is now building Conductor, described as “an app for managing teams of AI programmers.” The move from single-agent coding assistance (Melty) to multi-agent orchestration follows a natural progression in the AI coding space.

Technical Preferences

Elixir: His favorite programming language. He built ShlinkedIn (a satirical professional network) and StickerBaker in it.
Real-time systems: Many of his demos involve streaming—webcam capture, live narration, real-time image generation
Minimal dependencies: The Narrator project is 100 lines of Python

Key Takeaways

Ship publicly, learn quickly: Don’t hide demos until perfect. The feedback from Twitter posts is faster than any user research.
Combine primitives creatively: The Narrator isn’t novel technology—it’s GPT-4V + ElevenLabs + webcam. The creativity is in the combination.
Track git history in AI tools: Melty’s insight is that AI coding assistance should integrate with version control, not sit alongside it.
Start with fun: Holtz’s most successful projects are playful. Attenborough narrating your life is silly. That’s why it spread.