Key points
- TypeSafe AI's Diogo Almeida pitched Jev, released in September, as a fast, reliable model to call inside application code for automation.
- Glean uses Jev next to LLMs and other classifiers, and says it avoids repeated fine-tuning when its decision categories change.
- UiPath ranked document chunks in a tournament format with thousands of Jev calls, finishing in under a second.
- Workday prototyped a time-off card that begins filling in while the user is still typing.
Why it matters for marketers: Fast, cheap decision models could handle routing, ranking and classification steps in marketing software, leaving slower LLMs for writing and leaving teams with more predictable automation.
A session at the AI Engineer Code Summit focused on what the speakers call system one models. These return choices, scores or probabilities instead of generating text, and they tend to run faster and cheaper than reasoning LLMs. The showcase product was Jev, which TypeSafe AI released in September.
TypeSafe AI CEO Diogo Almeida argued that the core weakness of LLM-driven software is reliability, and that Jev calls belong directly inside application code. In his words, the goal is to automate everything. Glean's Chinmay Goyal said his company runs Jev alongside LLMs and other classifiers, and values that it adapts when decision categories change without repeated fine-tuning. He added that Glean's customers do not need to know which model is working behind the scenes.
UiPath engineers described a test in which Jev picked the best document chunks through rounds of small head-to-head groups, completing the job in under a second despite thousands of calls. A Workday principal engineer, Scott Roberts, showed an interface that starts filling in a time-off card as the user types, a shift he described as predicting actions rather than words.
For marketers and content teams, the pattern is relevant to tasks such as sorting content, ranking assets, tagging audiences and routing requests. These steps often do not need a full LLM, and a fast, consistent model could make them cheaper and easier to trust inside agent workflows.
Worth watching is whether teams adopt the layered approach shown here: small models for quick decisions, larger ones for generation. The speakers described the work as early, and Almeida compared the moment to the experimentation of the early internet.
Sources
This brief is our own summary. Read the full reporting at the source.