flagship · private · project
Tron
A decision model: probabilities, not text
A Jev-style decision model I run on my own hardware: send a state and typed questions - yes or no, one of several labels, or a score on a scale - and get probabilities back in about 20-300 ms instead of generated text.
The problem
Most of the decisions inside an agent are not prose. Is this true, which of these, how much: asking a chat model means generating text and then parsing a number back out of it, slowly and unreliably.
How it works
- An endpoint in the shape TypeSafe's Jev introduced ("System One" compatible): a state - a string or a set of named fields - goes in with typed questions, and probabilities come out.
- Three question types: the probability that a statement is true, a choice across labelled options, and a score on an ordered scale. Answers come back in roughly 20-300 ms, with up to 8,192 tokens of context per question.
- Measured before it is trusted: on its first real workload the base checkpoint was not reliable enough, so it runs in shadow while a fine-tuned model is trained on labelled data exported from that workload, and it earns a role only once it beats held-out tests.
Why it exists
The cheapest dependable answer to a yes-or-no question is a probability, not a paragraph. Tron is where I test that on my own workloads before anything depends on it.
More work
Hiring for AI or ML?
I am open to AI/ML Engineering, Data Science, and Python roles, plus research collaborations and consulting. New York based, shipping worldwide.