Charlie Holtz Builds AI Tools That Spark Joy

Table of content

Charlie Holtz builds AI tools fast. Really fast. As Hacker in Residence at Replicate, he shipped projects that hit #1 on Hacker News, made headlines in Business Insider, and even “struck fear into the heart of Hollywood.” His approach: build something fun in a weekend, share it publicly, and let the internet decide if it matters.

The Narrator Incident

In November 2023, Holtz posted a 45-second video that changed his career trajectory. He’d built Narrator, a Python script that uses GPT-4 Vision to watch your webcam and ElevenLabs to generate real-time David Attenborough-style commentary about whatever you’re doing.

“Here we have a fascinating specimen of the modern human, hunched over a glowing rectangle,” the AI intoned as Holtz sat at his desk.

The video went viral. Business Insider wrote about how Hollywood actors were worried about voice cloning. The repo gathered 4,400 stars. All from a script that takes about an hour to understand and maybe a day to build.

That’s Holtz’s whole thing: finding the fun intersection of available AI APIs and building something people actually want to play with.

Melty: Chat as Version Control

His most ambitious project is Melty, a code editor where every chat message becomes a git commit. The tagline is “AI code editor where every chat message is a git commit.”

The insight: existing AI coding tools create a mess. You end up with ten chat threads for one task, copy-pasting code between windows, and commits that mix human and AI work without clear boundaries.

Melty treats the conversation itself as the version history. Revert a chat message, revert the code. Branch a conversation, branch the work. The AI sees your git history and stays in sync like a pair programmer who actually knows what’s happening.

After 28 days of development, Holtz reported that Melty was writing about half of its own code. The repo has 5,400 stars.

The Project List

Holtz ships constantly:

CHARL-E — First Mac app for running Stable Diffusion locally
StickerBaker — AI sticker generator
EmojiGen — Text-to-emoji, hit #1 on Hacker News
ShlinkedIn — Satirical LinkedIn that made Der Spiegel’s front page
Jarley — AI that attends standup for you

Now he’s building Conductor, an app for managing teams of AI coding agents on your Mac.

The Stack

Holtz has strong opinions on tools:

Elixir — His favorite language, and he’ll tell you why
Replicate — For running ML models via API
GPT-4 Vision — For real-time image understanding
ElevenLabs — For voice synthesis

The pattern: use hosted APIs for the hard ML stuff, write the glue code in something enjoyable, ship before the weekend ends.

Background

Before AI, Holtz worked in quantitative finance at Point72. He studied at Brown, where he played ultimate frisbee with Jackson de Campos (his co-founder on Melty). He’s also 40 hours into pilot training, which explains the aviation weather quiz app (MetarQuiz) buried in his project list.

The finance-to-AI-prototyper pipeline is becoming common, but Holtz stands out because he builds things that make people smile. Narrator is technically simple but conceptually perfect—it shows what’s possible with three API calls and some good prompting.

The Philosophy

Holtz represents a specific school of AI development: build fast, build publicly, build fun things. Don’t wait for the perfect model or the complete feature set. Ship the minimum viable demo, put it on Twitter, and see what happens.

Sometimes nothing happens. Sometimes you accidentally terrify SAG-AFTRA.

His current project, Conductor, takes this further—instead of one AI writing code, manage a whole team of agents. It’s the logical next step for someone who’s been treating AI as a collaborator rather than a tool.