| Location: NJ / NYC (hybrid or remote preferred) Remote: Yes Willing to relocate: No What I’ve been working on lately: LLM introspection tooling and observability, mainly around token-level behavior during inference; Also experimented with inference-time interventions (no retraining) e.g. improved DeBERTa on HANS via targeted layer/head adjustments. I just put up a small demo of part of the tooling (HF Space): https://huggingface.co/spaces/anotheruserishere/Cartogemma It exposes: -per-head projections into token space (logit lens-style) -token rank changes across layers for target-tokens ("rank displacement") -top-k next-token branches with internal state views -mute a head at a given L x H coordinate -inject tokens or rewind context There's a minimal example in the UI showing how token candidates stabilize (or don’t) across layers Background:
~15 years in data science/ analytics (higher ed), mostly translating technical work into decisions & policy for leadership.
More recently focused on LLM internals + tooling (Python/Rust, local model stacks, etc.) and agentic analytical tools for operational use.
Background in philosophy / applied linguistics & NLP; previously designed and taught a course on propaganda (how language shapes reactions to media etc.). Looking for:
roles around LLM tooling, evals, interpretability, or applied AI where understanding model behavior is useful. Tech: Python, Rust, SQL, embeddings, local LLM infra Contact: jim.jdiv@gmail.com |