Today we're announcing Belvedir: the autonomous private AI platform. Belvedir makes it easy for startups and SMBs to autonomously train custom models and memory systems, host them privately, and continually improve them as they are used in production.
Belvedir is part of Y Combinator's S26 batch and currently in closed alpha, working with a few initial customers. We would love to chat with people interested in trying the platform out. Reach us at founders@fractalresearch.ai.
Here are some problems, and how Belvedir fixes them:
Models must be private
The world is shifting toward private intelligence as AI ownership and data sovereignty become bigger concerns. Belvedir runs private and local.
Models must be custom
Different people want their tasks done differently. And many tasks will never show up in any lab-accessible training set in enough volume for general models to learn them. Belvedir makes unique, custom models from your own data.
Learning needs to be brought back to the weights
A lot of products offer AI customization in the form of memory, graphs, or other kinds of non-parametric learning. This approach has good use cases, but it doesn't achieve the subtle interconnectedness of thought and idea fluency that comes with learning in weights. The best solution is a mix of both weight and memory updates. Belvedir is that mix.
Closed model inference is expensive
If you finetune an open-source model half the size of a closed-source one and give it a good harness, you can achieve SOTA at a fraction of the cost. Belvedir lets you finetune that model and make that harness.
Training models is expensive
Companies like Palantir and Applied Compute already offer private models to large enterprises that can afford expensive forward-deployed engineering. Small and medium startups, and certainly consumers, can't. Belvedir replaces that service layer with autonomy.
Current model training products are complicated
Most startup founders, and every ordinary consumer we've met, either have never used a model training platform or got lost in the ML jargon when they tried. We are building Belvedir to be simple enough for the average person to use intuitively.
Ambitions
We want to make private AI the default for any use case, starting with startups and SMBs, then consumers.
We are bullish on a world with billions, if not trillions, of small AI models. We want to build the ecosystem where these models are born, learn, and interact with each other.