Lab
The half of full-stack that is still moving. Projects where a model does part of the work, and I have to design around an answer I cannot fully predict.
Building on top of something that guesses
A model is not a function. It returns something plausible rather than something correct, and it returns it slowly and at a price. Every project here exists because I wanted to find out what that costs on the other side of the API - in the schema, in the interface, in the error states.
So the interesting part is rarely the prompt. It is what happens when the answer is wrong, empty, or forty seconds late, and a person is still sitting there waiting for the page to do something.
Generation
Text, images and audio produced on demand, plus the caching, retries and fallbacks that keep a page usable while that happens.
Structure from prose
Turning what a person typed into something a database can hold: schemas, survey logic, itineraries. The model proposes, the constraints dispose.
Interfaces for uncertainty
Progress that is honest about latency, output a person can correct, and a path back when the answer is simply wrong.
What is actually running
No single provider. Each project uses whatever answered the question best at the time, behind an interface thin enough to swap.
See every project- Google Gemini
- OpenAI API
- GPT-2
- Stable Diffusion
- Google Cloud Speech
- Agora SDK
- Django
- Flask