Augustin Chan

I'm an independent researcher. I published two papers on arXiv this year about multi-agent LLM behaviour, listed below. Since September I have been studying interpretability. My current question: do a model's own activations predict what it will do, such as breaking a promise, better than its stated reasoning does? Digital Rain Technologies is the umbrella the work ships under.

Before this I worked on enterprise AI architecture and on small, single-purpose software tools. I still build both. Read how this introduction came together.

I also produce and co-host the Hong Kong AI Podcast, a conversation series with practitioners building Hong Kong's AI future.

Previously, I was a Development Architect at Informatica, leading enterprise architecture for Fortune 500 customers — a global role with an APAC focus, with a lot of remote support for US customers. Before that, I was a Senior Consultant at Dun & Bradstreet.

I hold a B.S. in Cognitive Science with Specialization in Computation from UC San Diego.

Research

In progress

Do a model's activations predict when it breaks a promise?
A probing study on Qwen3.5-27B playing all seven states of the Warring States game against itself. It tests whether linear probes on the model's own activations beat a chain-of-thought monitor. The probe results are not in yet. A text-and-board baseline scores 0.53 to 0.61 AUROC across four games.
What changes inside the model when a prompt changes who wins?
In the Warring States paper below, I gave one agent (Han) a reflective framework prompt. Han's survival did not change, but the winner distribution across games did. Even scrambled text shifted it, and the framework's content did not predict any decision. Behaviour alone does not say how a small prompt change travels through a campaign. I want to use existing interpretability tools, or build new ones, to look at what the model represents. This runs on the same game engine as the paper. I have no results yet.

Published

All my links
linktr.ee/augchan42
Every project, profile and podcast in one place.
Latest
A Runner Self-Update Broke npm on All Five of My CI Runners
A GitHub Actions runner updated itself overnight and turned npm into a 54-byte file that could not find its own code. It hit all five self-hosted runners on epsilon at once. How Claude Code and I found it, fixed it and set up a 5-minute timer to repair it after the next update.Oct 2026

Essays

Short posts

Full archive

Building

Enterprise migration & readiness

  • Crawlopen-source
    Migration intelligence — automated discovery, risk scoring, and planning for legacy system modernization.
  • Graftopen-source
    AI-native BI report translation — parses Tableau, JasperReports, and FineReport into a vendor-neutral IR, then analyzes and translates report logic across platforms.
  • ARA Evalopen-source
    Framework for decomposing autonomous AI risk into 7 dimensions, so practitioners can evaluate whether agents can act without human approval.

Symbolic systems & classical computing

  • 8-Bit Oracle
    The I-Ching through the lens of retro computing and science fiction — tech-noir art and commentaries. Presented at Taipei Blockchain Week.
  • Six Lines
    iOS app for learning the I-Ching as literature and philosophy — bilingual hexagram reference with classical commentary and public domain masterwork art. Moving to sixlines.day.
  • Plum Blossom Computer (GitHub)
    Cosmological computation terminal — I-Ching divination (梅花易數) fused with on-chain commitments and real-time astronomical data. Built at the EasyA x Consensus HK 2026 hackathon.
  • seeded-iching-engine (GitHub)

Research platforms

  • QDayAnon Research Platform
    Built for bitcoinquantum.com and live in production — curated post-quantum research with agent workflows that publish end-to-end, to the newspaper front end and to X. It posted autonomously on my sandbox account, @theqi0, until I stopped paying for X API access; the client chose to keep the last step manual on their own.

Podcast

  • Hong Kong AI Podcast15 episodes
    I produce and co-host real conversations with AI practitioners building in Hong Kong — covering research, fintech, robotics, spatial computing, healthcare, culture, and the local AI scene.

Speaking

Client work