Find Your Summit
Earlier this year, Paul Conyngham, an engineer with no medical training, used ChatGPT to design a custom mRNA cancer vaccine for his dog.
Today, people work far outside their domains, and they do it well. Tools like Cursor, Claude Code, and Lovable let someone who has never shipped software to build real applications.
Shaping the model itself, however, is not democratized. Post-training is the work of a few hundred people at a few dozen companies. While everyone can use a frontier model, customizing one remains difficult.
Your edge is not the model
Conyngham's edge wasn't ChatGPT, but his dog's genome.
The differentiator is the data you own, and every time you run that data through a closed model, you hand your edge to someone else. Every correction, every rejection, every piece of context goes into improving something you will never own.
A general model is the same base camp for everyone. Everything above it is the climb.
The building blocks
Whitney AI makes post-training accessible.
The path to democratized post-training will be a long one as agents today can't one-shot post-training workflows. We believe this path starts with a few building blocks:
- Clustering: Your traces already contain every place the model got it wrong. Clustering surfaces the failures that repeat, so you're able to improve upon model behavior rather than one-off edge cases.
- Evaluations: Institutional knowledge lives in people's heads, systems of record, and production workflows. Encoding that knowledge into evaluations will give models a concrete metric to optimize for.
- Model training: Anyone can prompt a training loop. The hard part is knowing which method fits the failure you're fixing, and how to debug training runs when they fail.
- Model serving: By itself, training is just that. Inference is the invisible hand that brings models to users in a seamless and reliable manner.
- Continuous learning: A new model will start drifting the day you ship it. Routing, rollouts, and the traces coming back from real traffic will let the model keep improving after launch.
Why we're building it
We've each spent time in a different part of that loop: post-training research at Microsoft Superintelligence, applied AI and evals at YC startups, model serving at Netflix. Because we've worked the whole agent stack, we believe everyone should have access to making and improving models they own. This is why we're building the IDE for post-training, for AI engineers and teams who want a custom model, but don't have a research org to build one.
Intelligence belongs to the people who build with it.