About

I'm Tom Vaucourt. I like building things with code, but I never liked writing code for its own sake. At first the bottleneck was code itself. In prépa I skipped the SQL course because I was sure I'd never need it, and at Centrale I spent 20 to 30 hours clicking through a vendor's software to export scans for a hundred patients, because I couldn't script my way out.

Then it was compute. I rationed myself to a $20 ChatGPT plan for a year and a half, and bought a 3090 to run models at home. I learned prefill and the KV cache because the model had to fit in 24 GB.

Then the bottleneck was me. I wouldn't let an agent near my home lab. Four months ago I still read every line agents wrote, and said vibe coding wasn't for production. Once I let go of owning the code and kept only the decisions, my GitHub graph changed the same week. Until mid-June I had months with no commits at all. In September I made 1,590 commits and opened 702 pull requests.

Letting agents write the code didn't lower the bar. When localvoxtral started lagging on long dictations, I gave an agent the problem and the numbers to move. It profiled mlx-audio-swift, found the slow steps and wrote the fixes. I trust only the numbers at the end, and I read each PR as if a stranger had sent it, because that's how the maintainer will read it. Eight are merged. The first ones took a 60-second stream from 114.7 seconds to process down to 0.76 seconds per second of audio. The last three fixed hour-long sessions: live text used to fall behind the speaker after fifteen minutes, and now it keeps up for three hours with an identical transcript. I also have merged fixes in llama.cpp and CodexBar.

Now I run about 20 agents at once, and an orchestrator keeps them moving. It brings me two kinds of decisions. The first is logistics. localvoxtral is a Mac app and my MacBook is the only Mac I own, so jobs that need it wait until I'm asleep. I also set which model does what across my four subscriptions, because each is good at different things. The agents track the usage left themselves, through CodexBar. The second kind is taste. I make every decision a user can see, down to single UI elements, and I tell the agents how I actually use the app, because they never use it.

What's left for me is ideas and feedback. I built localvoxtral so I can talk to agents instead of typing, and its inbox lets me dictate a passing idea and have it filed as an issue in the right repo before I lose it. For feedback I need users. One of them keeps dictation on all day and never presses a key, and that changed what I think the app is for. It should never get in the way.

So feedback is my bottleneck. Tell me where I'm wrong: tom.vaucourt@gmail.com