Eric J Ma's Website

In this post, I introduce prompt cards: small cards, inspired by Model Cards, that teach one prompt I actually use. Each card numbers the parts of the prompt that matter, explains why each part matters when you hover over it, and names any skill or AGENTS.md instructions the prompt relies on. I explain why I think this works as a teaching tool for working with AI, walk through one card part by part, and show how you can make your own with my prompt-card skill. When you share a prompt, are you also sharing why it works?

In this post, I share the knowledge wiki pattern I've settled on in Obsidian: an index page, maps of content, a flat folder of raw sources, and evergreen notes, and nothing more. I walk through what each element does, how they link together for progressive reveal, and how I split the work with coding agents: they ingest sources, name files, and maintain links, while I keep control over themes and do the thinking that makes the notes worth having. If four moving parts can hold everything you're trying to learn, what are the rest of yours doing?

In this post, I walk through my setup for agentic data science with marimo notebooks: picking a harness and model, writing instruction files that hold the agent to my standards, sandboxing environments with uv or pixi, and pairing the agent with a live notebook in an iterative loop. I also cover custom anywidget viewers, feedback with marimo-lens, and turning the finished notebook into a presentation. It all started as a whiteboard sketch headed for a video and course. What could you discover if an agent typed while you thought out loud?

In this post, I recount live-demoing an hour-long reanalysis of my published IRED-88 enzyme data at UMass Chan, pairing a coding agent with a reactive marimo notebook. New tools marimo Lens and Studio let me point at plots and project slides straight from the notebook. The agent flew through charts but stumbled on the epistasis transform I knew from hard-won experience. It's a candid look at data science in the age of agents. When machines master the obvious moves, how do we cultivate the judgment that remains ours?

Laya is an open-source decision model that answers structured questions in a single forward pass, no text generation involved. I ran it through email triage, multilingual postmarket surveillance, real pharma document classification, a fine-tuning experiment, and a Modal deployment to see where it shines, where it breaks, and what its probabilities are really worth. Curious whether a 421M-parameter model can do the boring 90% of an AI pipeline cheaper than a giant LLM? Read on to find out!

In this post, I share how I forked Seems, a Python-flavored judgment language, into seems-laya, swapping the hosted judge for an open-weight model running on my own machine. I walk through the one-file swap, getting judgment syntax into plain Python imports and marimo notebooks, and the concurrency, calibration, namespace, and bytecode-cache surprises along the way, most of them fixed by my coding agent. I also wrestle honestly with whether the work was creative at all. If I can't tell the difference between interpolation and invention, how would any of us know when a machine crosses that line?

In this post, I share how I built Sage, a community-wide Discord chat agent, by simply wrapping a standard coding agent inside a persistent container. I'll walk you through the surprisingly straightforward architecture: a websocket, a subprocess, and a single hand-curated markdown file. We'll explore how multiplayer settings shift the rules around grounding, consistency, and cost, and why keeping the stack boring actually makes it more reliable. If you've ever wondered how to scale an AI assistant beyond a single user without overcomplicating things, what's the simplest way you could start building one for your own community?

In this post, I share what I learned at a conference on AI in GxP, the regulations keeping medicines and devices safe, arriving from the research side where the acronym meant nothing to me three months ago. I sketch a map of SOPs, validation, and inspections, and how stochastic AI systems force a rethink: defining a range of acceptable behavior instead of pass/fail checks. I close with what data scientists can offer, from model cards to evaluation harnesses. If regulators set the principles but leave the numbers to us, who gets to decide what "good enough" means?

In this post, I share my hands-on review of the PLAUD Note Pro, a credit-card-sized recorder that carried me through a two-day conference on AI in GxP. I describe how it captured twenty-odd talks, survived two days without a charger, and let me highlight key moments with a single button press. I also walk through feeding the transcripts and slide photos into my agent pipeline and Obsidian vault. If a device this simple can disappear into a workflow and quietly do one job well, what else might we be overcomplicating?

When my team was stuck talking past each other with no slam dunk option in sight, I had everyone bet 1,000 points of funny money across the options, and the allocations revealed the team's collective belief distribution. It beats voting because it captures preference strength and confidence separately. Curious how a 70-point bet and a 900-point bet unblocked the discussion?