praxagent / Work index
AI research and open-source software
Explore experiments on language models, the Prax agent tools, and focused applications. Each project links to its research, code, or documentation.
01 / Research programs
Model internals and evaluation
Layer similarity
How the fitted lens’s readouts relate across layers.
Model scale and layer structure
A comparison across sizes and model families, with architecture as a possible confound.
Response to internal edits
Linear edit delivery produces a downstream response that is not proportional to dose.
Detecting steering
A matched clean run reveals a steering signal, but related features can produce similar signals.
Jacobian Lenses & J-Space
Open-source tools for studying model representations, with controlled experiments and reproducible results.
Sparse Autoencoders
Feature activation mapping, label-validity tests, balanced contrasts, and reproducible analysis on open models.
2026 Internship
A remote mentorship in AI, biology, and technical writing. Position filled.
A self-hosted system for coordinating AI agents, with a shared workspace, sandboxed tools, plugins, and optional isolation of provider credentials.
GitHub ↗
Prax
Coordinate agents with persistent memory, tool permissions, and integrations for chat, code, and browsing.
TeamWork
A shared workspace for chat, task boards, files, terminal access, and browser interaction.
Prax Sandbox
A container-based environment for coding agents, browser automation, and desktop tools, with a Python client.
Prax Plugins
Self-contained plugin pack for presentations, music, radio, and other runtime extensions.
Secrets proxy
Keeps provider API keys outside the agent process when deployed as an isolated service.
03 / Focused software