In research · deployment aheadStasis
An AI worth spending time with. A mind that asks better questions.
Stasis is my family of models, and the center of my independent research. The ambition is personal: build the best AI to hang out with, with the curiosity and contemplative presence of a Greek philosopher.
I am currently creating philosophical, multi-turn conversations: exchanges that stay with an idea, question an assumption, and make room for a thought to develop. This is an active research direction; public deployment is still ahead.
Stasis Coder is the coding-focused work within the family. It explores fine-tuning, curated training data, debugging, and full-stack reasoning. The public repository documents this work; it is not a hosted model demo.
What makes an AI worth staying in conversation with?
Stasis explores two kinds of depth: the precision to build working software, and the curiosity to stay with a difficult idea.
A model family
Stasis is the broader research identity. Stasis Coder is its focused exploration of software engineering.
Conversation as craft
The current direction is philosophical, multi-turn dialogue that develops a thought across an exchange.
Independent research
Dataset design, fine-tuning experiments, and evaluation are part of one ongoing engineering practice.
A mind you want to spend time with.
The long-term goal is an AI companion with the contemplative presence of a Greek philosopher: curious, thoughtful, and willing to examine an assumption. The interesting unit is an entire conversation. Can an idea become more precise as the exchange continues? Can a question open up something worth thinking about?
That ambition shapes the current work on philosophical multi-turn conversations. Each exchange creates an opportunity to explore continuity, disagreement, follow-up questions, and the space a person needs to develop their own position. These are research goals; public deployment is still ahead.
Go deep enough to make the software work.
Stasis Coder concentrates on a specific full-stack workflow: React, Next.js, Tailwind CSS, and the connections to Node and Postgres-style backends. Its public research notes describe QLoRA and Unsloth as the fine-tuning approach, with training examples organized around realistic engineering tasks.
The dataset methodology emphasizes execution checks: linting, type checking, and headless rendering where applicable. Other categories cover UI descriptions, state-management edge cases, architecture choices, and debugging conversations that begin with an imprecise human report. The underlying idea is that plausible-looking code is only the beginning of the task.
Make the next turn better than the last.
The coding and conversational directions share a useful question: what happens after the first answer? A coding model has to respond to a failed build, a visual mismatch, or a changed requirement. A conversational model has to remember what matters, notice tension in an argument, and ask a relevant next question.
A useful evaluation would therefore look beyond isolated responses. For coding, that means working artifacts and successful corrections on held-out tasks. For philosophical dialogue, it means consistency, responsiveness, and the quality of the developing exchange. These are the criteria that make the research direction interesting, rather than a claim that they have already been solved.
How it comes together.
Define the behavior
Choose a concrete interaction: debug a broken UI or sustain a philosophical exchange.
Build the examples
Construct multi-turn data around the task, its context, and the next useful response.
Check the signal
Audit examples and validate executable code where applicable.
Train & evaluate
Explore specialization, measure held-out behavior, and use failures to guide the next iteration.
The conversation is the product.
Stasis connects model engineering with a personal question about the kind of intelligence worth building. The work is ongoing, with a public model release still to come.