The Mechanism¶
The Mechanism is a research cluster built from found infrastructure: an existing desktop and a disused Raspberry Pi cluster, with a few pieces added where those ran out. It began as an exercise in critical technical practice. The aim was to learn about the affordances and constraints of AI for humanities and social science research by engaging in system design and maintenance myself, not paying for online services where I would be one of many paying customers, or renting cloud infrastructure.
The result is unlikely to be implemented by anyone else. I made few concessions to simplicity or future users: it is a radically bespoke system, designed and configured to support my technical and scholarly practices. Some of the design principles and architecture might be useful to people designing AI enabled research platforms, but that would be incidental to my purpose.
I am only keeping The Mechanism in service because it has proven to be surprisingly useful for my research, exceeding my expectations for what is closer to a Heath Robinson machine than a refined computing system. I use The Mechanism on a daily basis, as the primary research tool for my next monograph, Infrastructural Inflections: Critical Technical Practice and the Architecture of Knowledge. The system is a working research system first, and an experimental research platform second.
If I was forced to provide a single justification for the contraption's continued existence I would point to the balance between a closed epistemic universe of scholarly sources and the power of persistent agentic AI: the balance is difficult to achieve but compelling. My suspicion is that AI-enabled humanities research requires a radical narrowing to high quality sources tailored to the individual researcher, coupled with persistent agentic AI that can guide, critique, and learn alongside the researcher. This requires a lot more than straight-forward "chat with my pdfs" functionality, but significantly less than access to the entire World Wide Web. The Mechanism keeps the researcher firmly at the centre of epistemological and methodological decision-making but with access to the full range of their own sources (primary, secondary, RSS, social media) along with a consistent 'mentor and guide' and a range of configurable 'critical friends'. By keeping as much processing offline as possible and retaining detailed records in offline Markdown files I feel in control: I could continue to work on my monograph project even in the worst case scenario where access to frontier AI is entirely disabled. Frontier AI services are only accessed intermittently, for advanced text processing and occasional 1-1 discussion sessions and workshops with multiple AI models. I expect to come across a service that renders The Mechanism redundant one day, but I would only need to upload some Markdown files and plug in my Zotero library to migrate to it.
This site documents the system at a technical level: what it is, how it works, what it costs, and what building it has taught me.
James Smithies, Australian National University. The Mechanism is funded by the ANU Futures Scheme (2024–2026), awarded when I joined the ANU.