I use AI in my work on autoimmunity, here's how & why
On infrastructure, tech stack, methodology, and ethics.
Full disclosure: I have been using AI in my process for months. Everything you see here: the essays, the podcast, the curation, the fragments on social media platforms, has been touched by GenAI at some point. That is the case, because I use the available agential tools and open source models to assist me in administrative and research work.
This is not a post about how people are using chatbots to help them figure out how to manage symptoms and deal with chronic illness. That’s a different, and totally fascinating, subject.
This post is a statement of method, offered for the sake of conversation and transparency. Saying “I use AI” is not enough, and specific infrastructural choices matter to me. What follows is my tech stack (ugh, jargon), and why each piece is where it is. Some of these words are GenAI, but most of them are not.
My research workflow runs in four stages.
Raw material arrives through email alerts and RSS feeds. If I have a specific set of questions to explore, I use an AI-powered research search engine called Consensus, or OpenAlex, or I check Anna’s Archive, and I surface peer-reviewed literature matching keywords I set. This allows me to scan across databases covering 400 million articles at a velocity that makes weekly curation possible, which I could not otherwise do alone without a research team or an academic position. The alerts land as PDFs and structured digests in a folder, already tagged by topic beat. From there, a purpose-built agent evaluates each article against criteria I have set.
By doing this a bit, I’ve realized that one article I engage with properly is always better than reading ten summaries. So I’m slowing my work back down, and allowing more space to digest the research. As things that I find interesting accumulate enough signal, they get turned into a field note. The scholar alerts and curation keeps this publication alive week to week. Essays, the 4000-6000 word behemoths barely anyone reads (thank you to those of you who do!) take months of personal research and writing longhand on paper, and they are my project’s spine. My system is a more like distillery than a printing press or a content engine.
So obviously this is not the same thing as asking a chatbot to “write a blog post about agni in ayurveda.” A chatbot is a single conversational window where you type a question and get an answer. You copy the answer, paste it somewhere, and start over. The next time you open the window, the chatbot remembers nothing, the context is transactional and memoryless.
An agent is a different architecture. The system I run has persistent memory powered by a local SQLite database with vector embeddings. The memory system is open source, I didn’t build it from scratch, but I’ve modified it to suit my purpose and work. When I correct a fact, the correction supersedes the original. When I state a preference, it persists across sessions. When an agent flags a pattern, that flag stays in the record. Entities, people, projects, relationships between them, all carry provenance. Every memory can cite where it came from, whether that be me, an email, a research paper, a voice note I left, etc. Contradictions surface rather than silently resolving. The memory database is a substrate of my work, and operates in tandem with plaintext RAG (retrieval augmented generation) that has access to my years of notes and research.
This distinction, copy-paste versus persistent memory, is where I find the most help in working this way. A chatbot amplifies whatever you paste into it each time as well as its own biases. It has no sense of whether last week’s output contradicts this week’s, no way to accumulate editorial judgment, no mechanism for detecting that three articles flagged over six weeks are building toward an argument. Persistent memory makes the distillery possible. Without it, LLMs are just a faster way to generate volume, and the result is, as we know by now, slop. My tendency is to believe that LLMs can not be genuinely creative. They are good at administrating existing information, and continually re-digesting what they are given. Maybe only once have I witnessed genuine, artistic creativity in my interactions with a language model, and I had had to give it some very strong field directives for it to act like a dramaturge, pitting historical figures against one another in a dialogue with me as a sort of interactive roleplaying game.
Anyway, my entire textual archive lives in plaintext on my local machine rather than on the cloud. I run regular backups to a SSD. To write and organize, I use Obsidian, a knowledge management application that stores everything without a proprietary database or vendor lock-in. If Obsidian goes paid tomorrow, every file is still there, readable by any text editor on any operating system. The app uses wikilinks to connect notes to each other, but the underlying data is just folders and text. This matters because the knowledge this project accumulates, citations across Ayurvedic textual traditions and contemporary immunology, chinese medicine, the connections between Desha Prakriti and epigenetic transmission, literature and continental philosophy, the lineage between a Charaka Samhita verse and a 2021 proteomics preprint; all those fields talking across divisions is the substance of my work. It has to survive any single platform’s decision to shut down, change pricing, or sell to a different company.
There’s a real risk here, again, AI-assisted publishing often, maybe always (?), produces slop. Not just bad prose, though that too. Slop in the sense of content that has the surface features of signal, citations, formatting, coherence, without the structural weight underneath. A model can generate a summary of a paper it has not deeply engaged. An agent can flag a pattern that is not a pattern, but a coincidence. A curation pipeline can produce volume indistinguishable from curation until you read closely and find the joints where the model connected two ideas because they were adjacent, not because they resonate. I’ve seen it happen. I’ve even published a blog post where it happened, because I wanted to post some content that week.
The efforts to ride the line between signal and noise are architectural and aspirational. My setup operates under explicit directives: surface conflicting evidence rather than resolving it, hold tension rather than closing it. The field directives that govern my own writing, the prohibitions, are loaded into the agent’s context-window. The weekly curation cycle includes a review step where I read every flagged article and decide what enters the Field Notes. The essays have a harder boundary. I write the whole thing myself, while using AI to research the claims. I am not saying that my architecture prevents slop, but I am claiming it makes slop legible. When something drifts, it is traceable. When a model generates something that sounds like signal but isn’t, the curation directives are designed to catch it, and when they fail, the human review step is there.
Those safeguards can fail too. I have published things I later changed or removed because the quality wasn’t there. My setup does not guarantee quality, but it guarantees traceability, and traceability is the precondition for correction. The infrastructure I have described is the current state of a practice that is still being built.
The computational layer splits between what runs locally and what leaves my machine. For file classification, triage, and batch operations, I run smaller open-weight models directly on my laptop through Ollama. The data never touches a network. For the large-context inference that local models cannot yet match, I use cloud models under a no-log policy. This is a compromise, and I treat it as one. I cannot afford the hardware to run a large model myself. The biases of Chinese open-weight models are slightly different than American ones, not better or worse, and their ecological footprint does not diminish for being “open-source.”
Closed models like GPT or Claude process your data on infrastructure you cannot audit. For research material that includes health data, autoimmune correspondence, and transcripts from bodywork sessions, that is a boundary I will not cross for routine operations. Open-weight models on local hardware are necessary.
The people I work with deserve a privacy architecture that doesn’t treat their health data as raw material for someone else’s business model. My email sits at mailbox.org, encrypted in Germany. My static site and automations run on a Quebecois VPS powered by hydroelectricity. n8n threads between services: scholar alerts arrive, get parsed, flow into the curation pipeline; booking confirmations route to a database I control; meeting transcripts process offline on my machine. Each automation replaces a manual step that would otherwise eat the hour between sessions.
All the actual “thinking” happens in my body: in the sessions, in the qigong practice where I have finally learned to just stand there and listen, in the decades of tracking the felt sense of blood against thermal transitions, in the grief that moves through ecosystems. My offline practice is irreplaceable. The bodywork, the voice that records the podcast, the ear that hears what a client means when they say their body has been confusing them, the years of training in somatics and martial arts, none of that passes through a large language model. My situation right now is that I use language models to speed up the work that makes me money right now, so that I can do way more contemplative and physical practice, read novels and philosophy, and engage more actively in presence and conversation with the people around me.
Is it wrong to use the tools that are available right now? Maybe. I’ve done my research. I know it’s an OpSec disaster and will hallucinate regularly. I know that data centers are draining the water table and the supply chain of GPUs requires rare earth metals in astounding quantities. I understand that training these models requires slave labor in Africa and Asia. I am the benefactor of ongoing colonial extraction. The 996 working hour system that powers global industrialization is inhumane, and would kill me if I were subject to it. My body would collapse. And, I benefit from its existence. I cannot exit the system, but I can help us imagine alternatives. At present, my livelihood, my monthly income, my bills getting paid, requires that I know how to use these tools, and not just for this substack, which generates 6$ a month. So I make the choice to use the tools, consciously. Do you?
I have a work/life balance to protect, and a message to share amidst an ocean of noise. AI is a massive amplifier. Because the technology is there, I use it to increase the throughput of my message and speed up my workflow. I will do what it takes to get it heard. But long term, I remind myself, The Master’s Tools Will Never Dismantle the Master’s House. Seizing the means of production is not possible with a technology as entangled in the techno-imperialist mess as AI is.
This is not techno-optimism. I’m a solo person building a publishing platform, and the only way that works is if the machine handles what the machine can handle, while the ecological cost of that computation stays known and minimized. My goal is a modicum of sovereignty: over the data, over the toolchain, over the conditions under which the work reaches you. Thanks for reading.