AI Engineer/Boston, Massachusetts
Brad Egan
I build AI systems and the tooling that measures whether they actually work.
01About
Most of what I do lands in one of two places: models I merge and evaluate, or tooling for the coding agents that have quietly changed how software gets written.
I came to this from philosophy: a BA at Trinity, then a master's in computer science at Northeastern concentrating in AI and data science. In 2025 I started EganAI, which is where the client work lives.
Consulting, custom models, and agent evaluation run through the company. egan.ai
- Based
- Boston, MA
- Company
- EganAI
- Studied
- MS Computer Science, Northeastern
- Before
- BA Philosophy, Trinity
02Open source
github.com/bradAGI →- awesome-cli-coding-agents1,021 ★Curated directory of terminal-native AI coding agents and the harnesses that orchestrate them.
- GraphMemory156 ★GraphRAG database with a hybrid graph and vector store.
- LLM-Shortcut25 ★iOS shortcut that reaches an LLM through a Docker endpoint.
- DSPy-Stock-News-Sentiment-Analyzer23 ★Sentiment analysis over stock news, built as a single DSPy notebook.
- awesome-free-inference13 ★Guide to free LLM inference APIs. Rate limits shift constantly, so the tables are a snapshot rather than a live feed.
- ultimate-gemini-skill8 ★Standalone Claude Code skill wrapping Gemini image generation.
03Research
huggingface.co/EganAI →Merges
- qwen3.5-9b-terminal-merge360 dl / 17 ♥Merge tuned for terminal and coding-agent work.
- Qwen3-4B-Instruct merge21 dl / 7 ♥Instruct-side merge produced with mergekit. Q8_0 GGUF quant alongside it.
- Qwen3-4B-Thinking merge11 dl / 4 ♥Reasoning-side merge from the same run. Q8_0 GGUF quant alongside it.
Experiments
- promptlabTestbed for comparing evolutionary prompt-optimization techniques head to head.Every technique runs on the same data under the same budget, so the only difference is how it searches, and the README keeps the washed-out results next to the ones that held.
- terminal-model-evolverEvolutionary model merging, scored against Terminal-Bench 2.0 with runs executed on Modal.Terminal-Bench is public, so a score here is measured on ground other people already run rather than on a harness built alongside the thing it measures.
- rss-graphKnowledge graph over a personal FreshRSS archive, running on a Raspberry Pi.The archive is already a citation network, so questions about coverage stop being heuristics and become graph queries: who reported a story first, and which claims independent feeds corroborate.
- agent-watchReads a coding agent's transcript at the end of each turn and rules it finished, stalled, or blocked.An agent that stops early cannot notice it stopped early: the turn is over, so there is no next call in which to catch it. The checks are bookkeeping, not judgment, like whether it claimed done while the last test run failed.
- LatticeLocal-first research workspace over arXiv.A local model pulls claims out of arXiv papers into a SQLite knowledge graph, and each claim keeps a verbatim evidence span, so an answer traces back to the sentence behind it.
- Open SocratesTechnology forecasting.It turns dated, source-backed evidence into traceable claims and append-only forecast revisions, then scores them once the question resolves.