pauljump Claim this profile

12 composable backend packages, 447 tests — search, analytics, job queues, payments, notifications, LLM integration. Extracted from 16 production apps. All SQLite-backed.

Persona simulation engine — structured 8-section cognitive profiles that model how someone thinks, not what they sound like. Source-grounded, predictive, schema-first.

Byte-level 16MB language model — the only byte-level submission in 622+ OpenAI Parameter Golf entries. Built the right size, not shrunk from the wrong one.

CLI that solves AI coding agent amnesia. Persistent knowledge store (FTS5 SQLite) + automatic session injection = compounding returns on every hour invested.

Open tools to pull every NYC apartment's real price and history out of the walled gardens, plus a standardized public record of every building. Point your AI agent at it and build. The rent is price-fixed; we unfix it.

A reproducible pilot for retrieval, verification, and claim-bounded evaluation in legal AI

Applied AI systems case studies by Paul Jump

A benchmark built to evaluate and improve agent capabilities for supporting legal work.

Back up your X (Twitter) bookmarks to enriched JSONL — full tweets, threads, and article bodies. Real browser + your own cookies. No API, no LLM, $0.

Jam on a walk, ship real work. A voice call with an EM who runs a team of AI engineers on your own Mac.

A tool built to understand municipal permitting: a live AI research agent + a sourced ledger of what happens to permission after the permit.

Mine your own judgment from your AI transcripts, then measure how much of you a model actually carries. Not a mirror, a whetstone.

Turn the Wayback Machine into a structured, fully-sourced dataset of any one kind of thing — read by a local LLM (MLX), for free, on your own machine. Every fact carries its source.

Public review carve-out of Kit's care-graph package. Not open-source.