Hey, I’m Zachary Yu, the founder of Belvedir. It’s nice to meet you!

TD;LR: 

Private AI models are cheaper, better, faster, and protect your data. But right now they’re only used by big companies since they’re really difficult and expensive to make. Belvedir makes private AI models autonomously with less than 5 minutes of set up. By making it cheap and easy to make models, we hope people use Belvedir to create trillions of them.

Right now, we’re live with 10 companies who are seeing up to 10x lower inference costs and 2x improved performance on benchmarks. 

https://youtu.be/MXMfsqLFr3g 

How it works:

1. Get an API key from Belvedir and install our SDK
2. Belvedir turns your traces and logs into organized training sets and RL environments
3. We autonomously train models from your data and optimize harnesses on top of these models
4. We make and run benchmarks on these new models and harnesses
5. We deploy these models and give you an API key to access them through 
6. We give you a router that gradually send more traffic to your new model as it out performs your old one
7. As you use the new model we collect more traces and logs
8. The loop repeats

Your data is only used for you. Your models are only accessible by you. We use LoRA and can charge by the token by hosting multiple adapters on one cluster (so no worries if you have spikey traffic). 

About Me:

I’m Zachary Yu, the founder of Belvedir. Before this, I founded Traverse: a YC startup that sells RL environments to frontier labs, studied CS at UWaterloo, and made RL environments at Mercor.

Long Term Ambitions:

We want to make private AI the default for every use case. Starting with startups / SMBs and then consumers. 

We are bullish on a world where there are trillions of private AI models. We want to build the ecosystem where these models are made, hosted, and interact with each other.

Asks:

1. If you have a high inference bill, want better performance from your agent, or want to protect your data from frontier labs, we can help you out.
2. We are looking to hire engineers who have worked on RL / optimization, model inference, training infra, or just have that dog in them

Contact us at founders@fractalresearch.ai