Skip to content
View nmcbride-42's full-sized avatar

Block or report nmcbride-42

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
nmcbride-42/README.md

Hey, I'm Nick 🦍

I build AI agents. At work that means a GenAI agent platform at an electric utility: it answers questions from over a million pages of documents, shows where every answer came from, and can take action in ServiceNow. It also has to pass NERC CIP, which keeps you honest.

At home it means Iris, my own persistent agent. She has her own memory, runs around the clock across a few machines at my house and a VPS, and coordinates a crew of other agents when there's a lot to build. My favorite part is that we measure her own pieces and rip out the ones that lose to something simpler. Half of what I learned about agents came from the stuff we deleted.

Things worth a look

  • iris-systems is how Iris is put together: layered memory, hybrid search, hooks that enforce the rules instead of hoping the model remembers them, and the parts we retired after they lost to a baseline.
  • hallmark is my answer to "how do you stop an LLM from grading its own homework?" The model proposes, a verifier it can't touch decides, and "not sure" counts as a no.
  • touchstone: LLMs write programs, a sandbox runs them, and exact scorers decide what's real. It found a 111-number set in 0 to 999 with no three evenly spaced, beating the standard 105.
  • orrery is a small graph network that learned particle physics from a simulator and runs a world forward on my gaming GPU.
  • magpie digs through NASA TESS data for variable stars nobody has catalogued yet, with LLMs judging the light curves and the plots.
  • claude-code-guardrails is a Claude Code hook that won't let your agent wipe out uncommitted work. I built it after mine did.
  • lexicon is an experiment we wrote the rules for up front. The answer came back "no", and I kept it anyway.

How I build

I'm a designer by training and an architect by trade. Most of this was built pair-programming with Iris and Claude: I bring the architecture and the judgment, they do a lot of the typing, and nothing counts until a test or a measurement says so.

Off the keyboard

Hunting variable stars, metal detecting and rockhounding around eastern Washington, and keeping a 3D printer busy.

Say hi on LinkedIn

Pinned Loading

  1. claude-code-guardrails claude-code-guardrails Public

    PreToolUse safety hook for Claude Code: backs up dirty git trees before destructive ops, blocks unrecoverable ones.

    Python

  2. hallmark hallmark Public

    Fail-closed verification for LLM-proposed science: predeclared, hash-pinned verifiers decide on frozen TESS data

    Python

  3. iris-systems iris-systems Public

    Architecture of Iris, a persistent multi-model AI agent: layered memory, hybrid retrieval, hooks-first harness, measured retirements

    Python

  4. magpie magpie Public

    Tiered TESS light-curve funnel: numpy filters, catalog + common-mode vetoes, hostile LLM text/vision judges, VSX packets

    Python

  5. orrery orrery Public

    Neural particle world engine: a ~1.6M-param graph network learns physics from classical simulators and rolls worlds forward

    Python

  6. touchstone touchstone Public

    Generate -> verify -> artifact engine: LLM proposers write code, sandboxed exact/held-out scorers decide what counts.

    Python