Resume/CV/Experience: working demonstrations available below, anything else available on request.
Email: jim.jdiv@gmail.com
Seeking: applied AI, model behavior, evaluation, interpretability, inspectability, research engineering, synthetic data generation, or adjacent roles
I have spent more than 15 years building my own tools when the available ones were insufficient, and being someone sent when there was a problem, either to figure out what was going on or solve it, or both.
This has been as a generalist across analytics and data science for the operational arm of a large public university w/ steadily expanding domain responsibility, informing senior leadership, at times authoring strategy and policy. (I'm not as faculty, although I developed and taught a course for several years, Language of Propaganda, as an adjunct lecturer: adversarial uses of language examined through informal logic, cognitive blind spots, and case studies.) Separate from this, when a self-funding hobby found unexpected product-market fit I grew it to side-business and shipped 10,000+ items.
My academic background is in applied linguistics, NLP, cognitive science, and analytic philosophy. I might have followed a research or academic path, but by the time I completed my master's degree, I had concluded that the paths then available would not give me much room to pursue the questions I actually cared about.
Those interests are language, mind, cognition, and computation, all now converged in modern AI, I have devoted a lot of time bringing in ideas from traditional linguistics and some other areas and turning them into practical tools.
Recent work, including live demos:
- Cartogemma: A REPL-like environment for LLM inference. Explore generation as a branching process, preview output, inject tokens, rewind, ablate or restore heads, trace a chosen token through the layers. Per-head projections, residual contributions, the ordinary logit lens, and full-layer output side by side. The CMD/REPL bar functions once loaded. https://huggingface.co/spaces/anotheruserishere/Cartogemma
- Tokescope: watch a model in real time during inference, flag tokens to monitor & intervene. Catch something surfacing, before output. Once flagged it either logs it, stops the response with a hard gate, or suppresses using a gram-schmidt projection that takes the direction out of the vector. https://huggingface.co/spaces/anotheruserishere/Tokescope
- An instrumentation and intervention library for examining model internals during inference & using them to decompose behavior and outputs. It's what provides core tooling for the HF spaces listed. It analyzes derived structure, across layers and heads. These traces then use linear-algebraic, statistical, and overlap methods such as SVD, PCA, correlation analysis, and Jaccard similarity, results of which can then be used to steer a model through targeted activation-space interventions.
- Another library builds off of this one in the direction of mechanistic interpretability, for finding SAE-like features without the hassle of training an SAE, providing a range of static and interactive visualizations, scanning & storing model states at some or all steps of inference, among other things.
If any of this maps onto a problem your organization has or a role for which you have not found an easy title I would be glad to talk: jim.jdiv@gmail.com
I've had multiple people reach out telling me I buried a significant detail in my post: developing the curriculum on the adversarial use of language, and teaching it for a few years as an adjunct professor. So, here it is, in detail.
To be clear, I did not create the course itself. I had discretion on the specific material and approach to the course so long as it fit appropriate catalogue requirements: a multidisciplinary offering on the use of language to manipulate & persuade.
My own implementation had 3 pillars, with a central, loose thesis: Fundamentally, propaganda consists of poor arguments made in some way incomplete, incorrect, or insincerely, and simply learning to recognize their hallmarks is a 1st line defense against it. The first pillar used an excellent book, Logical Self Defense, as the foundation and I tried to entertain as best I could while hammering home its tenets on logical fallacies, premises, sufficient claim support, etc. The second pillar was a little light psychology and cognitive science. Influence techniques, psychological studies, cognitive blind spots, Benjamin Libet's readiness & awareness experiments, eye tracking saccades, even a live demonstration on the ease of implanting false memories through word choice. The message being that the human mind isn't perfect, it takes shortcuts, and care & awareness of that fact are the minimum requirements to avoiding some of it. The third pillar was simply case studies in propaganda. Readings, movies, news stories, as long as it didn't derail conversation I changed things to whatever was most topical at the time.
Resume/CV/Experience: working demonstrations available below, anything else available on request.
Email: jim.jdiv@gmail.com
Seeking: applied AI, model behavior, evaluation, interpretability, inspectability, research engineering, synthetic data generation, or adjacent roles
I have spent more than 15 years building my own tools when the available ones were insufficient, and being someone sent when there was a problem, either to figure out what was going on or solve it, or both.
This has been as a generalist across analytics and data science for the operational arm of a large public university w/ steadily expanding domain responsibility, informing senior leadership, at times authoring strategy and policy. (I'm not as faculty, although I developed and taught a course for several years, Language of Propaganda, as an adjunct lecturer: adversarial uses of language examined through informal logic, cognitive blind spots, and case studies.) Separate from this, when a self-funding hobby found unexpected product-market fit I grew it to side-business and shipped 10,000+ items.
My academic background is in applied linguistics, NLP, cognitive science, and analytic philosophy. I might have followed a research or academic path, but by the time I completed my master's degree, I had concluded that the paths then available would not give me much room to pursue the questions I actually cared about.
Those interests are language, mind, cognition, and computation, all now converged in modern AI, I have devoted a lot of time bringing in ideas from traditional linguistics and some other areas and turning them into practical tools.
Recent work, including live demos:
- Cartogemma: A REPL-like environment for LLM inference. Explore generation as a branching process, preview output, inject tokens, rewind, ablate or restore heads, trace a chosen token through the layers. Per-head projections, residual contributions, the ordinary logit lens, and full-layer output side by side. The CMD/REPL bar functions once loaded. https://huggingface.co/spaces/anotheruserishere/Cartogemma
- Tokescope: watch a model in real time during inference, flag tokens to monitor & intervene. Catch something surfacing, before output. Once flagged it either logs it, stops the response with a hard gate, or suppresses using a gram-schmidt projection that takes the direction out of the vector. https://huggingface.co/spaces/anotheruserishere/Tokescope
- Bertographer is similar, runs on encoder models, classification task, ie NLI models, Also with ad-hoc steering https://huggingface.co/spaces/anotheruserishere/Bertographer
- An instrumentation and intervention library for examining model internals during inference & using them to decompose behavior and outputs. It's what provides core tooling for the HF spaces listed. It analyzes derived structure, across layers and heads. These traces then use linear-algebraic, statistical, and overlap methods such as SVD, PCA, correlation analysis, and Jaccard similarity, results of which can then be used to steer a model through targeted activation-space interventions.
- Another library builds off of this one in the direction of mechanistic interpretability, for finding SAE-like features without the hassle of training an SAE, providing a range of static and interactive visualizations, scanning & storing model states at some or all steps of inference, among other things.
If any of this maps onto a problem your organization has or a role for which you have not found an easy title I would be glad to talk: jim.jdiv@gmail.com