Zachary Gray

Jev is TypeSafe's model that answers questions with probabilities instead of writing text. I'm testing where it fits inside an agent loop.

The talk

I gave it at TU's AIML Graduate Research Showcase in October 2026.

Saving the Planet, One Token at a Time

I showed off this research at TU's showcase, in a talk about using Jev as two filters that protect an agent's context. One keeps what's relevant, and the other catches prompt injection. In my tests, the agent processed about half the tokens, and none of the planted attacks got through.

See the deck

The paper

It's under review, so for now this is the title and what it covers.

System One Tool-Output Screening for LLM Agents for Safe and Efficient Outcomes

Jev screens what an agent's tools return, like file reads and shell output, before that output enters the agent's context. One filter keeps what's relevant, and the other hides prompt injection. The paper tests both across several low-cost agent models. It's under review.

The code

Both repos are public on GitHub.

  1. The paper's code

    jev-context-filter-harness on GitHub

    This is the code behind the filters, where Jev screens what an agent reads for relevance and for prompt injection. It has the harness, the test beds and the run records.

  2. The plugin

    jev-claude-plugin on GitHub

    A Claude Code plugin that puts Jev around the agent's tool calls. Both filters from the paper already run in it, and I'm adding more. It's a work in progress and it changes often. I built it on Claude Code instead of starting a new harness, because people don't leave the agent tools they already love, and I'm one of them.

Back to research
Sit at the desk, in 3D(the 3D version)There's a 3D version of this site. It needs a bigger screen than this one.This browser can't run WebGL, so it stays flat.It moves a lot, and your system asked for less motion.