Skip to content
View marliechorgan's full-sized avatar

Highlights

  • Pro

Block or report marliechorgan

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
marliechorgan/README.md

Charlie Morgan

AI deployment engineer, London. I find the problem worth solving inside a business, build the AI system for it, deploy it into their real environment and make it reliable enough to hand real work. I also build the guards and evals for the agents I run.

Safety and evals

  • Safety guards for Claude Code: hooks that stop a Claude Code agent deleting files outside its folder, rewriting its own settings, sending data off the machine, destroying git work or exposing servers. They work out what a command would do rather than matching its text. 625 test cases, a 95-case evasion probe with 0 holes and 0 over-blocks, CI on macOS and Linux.
  • alignment-faking-inspect: an Inspect port of Greenblatt et al. 2024, with the seven published prompt paraphrases as a contamination check. The first run found no robust gap on Claude Sonnet 4.6 or Opus 4.8; I am re-scoring it after an outside report of a possible scoring issue. Listed in the inspect_evals register.
  • Merged upstream: Inspect #4311 (epoch reducers converted dict-valued scores twice), ControlArena #891 (rescoring one .eval file overwrote the output folder; fix and 3 tests), LangChain OpenWiki #881 (a symlinked evidence file aborted the whole update).

Building

  • Defyner, Founding Brand Engineer: multi-agent systems built and deployed inside clients' own infrastructure. The client work is private.
  • claude-setup: the Claude Code skills and agents I use for real work (briefing sub-agents, review, browser automation, writing), each with tests or evals.
  • Pâlons, a voice tutor for Jèrriais, an endangered language: Best Working Prototype, Digital Jersey AI Hackathon, 4 Oct 2026.
  • token-usage-graph: counts and prices the token logs every AI coding tool leaves on your machine.

Before

  • Helix Technologies (acquired by Staynex): led technical development of a multi-agent travel platform.
  • Best Real-World Solution, Digital Jersey AI Hackathon, Nov 2025 (one of three awards across 20 teams): a RAG tool for ADHD care.
  • UK patent application GB2607064.9 (pending): acoustic monitoring where raw audio never leaves the device.
  • MSc Audio and Music Technology, York · BSc Mechanical Engineering, Newcastle.

Also a DJ and producer (Toasted) and founder of the label DRUMGL, with tracks played on BBC Radio 1 and at Glastonbury and Boiler Room.

My goal is to help Jersey and the UK lead as the agentic economy arrives. Open to conversations about AI deployment in high-stakes domains, multi-agent systems, and safety tooling for agents.

LinkedIn

Pinned Loading

  1. alignment-faking-inspect alignment-faking-inspect Public

    Inspect AI implementation of the Greenblatt et al. 2024 alignment-faking evaluation

    Python 1

  2. UKGovernmentBEIS/inspect_ai UKGovernmentBEIS/inspect_ai Public

    Inspect: A framework for large language model evaluations

    Python 3.2k 906

  3. UKGovernmentBEIS/inspect_evals UKGovernmentBEIS/inspect_evals Public

    Collection of evals for Inspect AI

    Python 735 478