Models, agents, and the systems around them
Research
Tensor Lab investigates how AI systems work, where they fail, and what changes when people give them access to real tools, data, and infrastructure.
Our work spans open models, agent behavior, local inference, and practical tools for working with documents and conversations. We build working systems, test their assumptions, and share code that others can inspect and use. Human judgment stays central: a useful system gives people a way to check its output and correct it.
Explore our GitHub organization · Work with us
At BSides Atlanta 2026#
Saturday, October 3, 2026 · Georgia Tech Hotel & Conference Center · Atlanta
James Widner and Wes Widner bring two sides of our research to BSides Atlanta: what happens when AI represents a person, and what happens when a model’s safeguards can be changed. Both talks are in the Grand Ballroom.
· Grand Ballroom
Taking the Day Off: AI Stand-Ins for Zoom
James explores an AI stand-in that follows a live meeting, decides when to respond, and sends synthetic speech and video back into the call. Using his own voice and mannerisms, the prototype examines the boundary between a convincing digital presence and the moments that reveal the machine.
Session details
· Grand Ballroom
Unlocking AI: Breaking the Safeguards in Open-Weight Models
Wes examines how modifying an open model can change its refusal behavior, what that costs in accuracy, and what remains of the original safety assumptions once someone else controls the weights. The research asks how we should evaluate a model after it has been altered.
Session detailsTimes are local to Atlanta (EDT). See the conference schedule for the latest details.
From the lab#
Our public repositories turn these interests into tools and experiments. Here is a cross-section of the work.
Conversations, processed on your device#
Tab Recorder captures browser-tab and microphone audio, saves recordings locally, and runs Whisper transcription on the device. Optional summaries use Chrome’s built-in Gemini Nano when available. Recordings and transcripts stay on your machine, giving you source material you can revisit and check.
Documents you can search and inspect#
Text Extractor brings PDF and image text extraction into the browser, with on-device OCR and an interface for reviewing text alongside the original page. Page the Ripper connects USB and network scanners to a self-hosted service that cleans up scans and produces searchable PDFs. Together, they address the unglamorous work of making source material usable.
Agentic development you can practice#
Our Agentic Development demos make agent workflows concrete: clarify the task, build project instructions, write a failing test, divide independent work, and keep context focused. Each exercise is small enough to run, inspect, and discuss. They support our hands-on training in using coding agents with clear scope and verifiable results.
Infrastructure for running AI yourself#
We maintain deployment packaging for Edgen and Kokoros, including Helm charts and CUDA container pipelines for inference and speech synthesis. The organization also hosts a fork of the RTX PRO 6000 LLM wiki, an upstream resource covering deployment, quantization, and performance on PCIe GPU systems. This infrastructure helps make local experiments practical and repeatable.
Questions driving the work#
What are we actually trusting? A model’s behavior can change with its weights, context, and tools. We investigate model integrity, adversarial behavior, and the assumptions behind giving agents access to real systems.
What can people verify? A transcript can be checked against a recording. Extracted text can be compared with a page. A code change can be tested. We look for workflows where evidence remains available and people can catch mistakes.
What can run under your control? Local inference and self-hosted tools let researchers examine the models, data paths, and operational trade-offs directly. Our public software provides ways to explore those choices.
Work with us#
We welcome researchers, engineers, students, and domain experts who want to test these questions with us. Try a project, bring a failure case, contribute a reproducible experiment, or talk with us about a research collaboration.