Eric Ma

How to Do Agentic Data Science written by Eric J. Ma on 2026-02-01 | tags: agentic coding experiments logging reports journal plots iteration structure exploration

Agentic coding is transforming data science. AI is a collaborator, not just a tool. I distilled 10 lessons from real ML projects—like agent journals and prescriptive prompts. These tips saved me days and sparked new ideas. Which tip would you try? bit.ly/409K82l #datascience #aiagents

📣 @ericmjl.bsky.social is teaching an "Agentic Data Science" workshop to raise funds for #SciPy2026

🗓️ March 27 & April 3, 2-4 PM EST
🎟️ $1,000 minimum donation ($200 for students)

👉️ Sign up at agent-assisted-data-science.vercel.app
👉️ Pay to NumFOCUS (tax-deductible) at givebutter.com/RBJuGI.

Model feel, fast tests, and AI coding that stays in flow written by Eric J. Ma on 2026-01-25 | tags: llm autonomy supervision personality verbosity harness refactoring workflow testing ergonomics

Model 'feel' can make or break your coding flow. Harness and feedback loops often matter more than model scores. Switching models and adding fast tests changed my workflow for the better. Agree? Repost or share your experience! buff.ly/Nv2cC1k #ai #devlife

Sign up here! buff.ly/P1T2mh3

5/5

Video recordings available if you can't attend live.

If you've been curious about agentic data science but didn't know where to start, this gives you techniques you can use the next day.

And you'll be supporting the SciPy community at the same time.

4/n

The workshop covers:

- Prescriptive prompting for experiments
- Project structure for agents
- Logging and verification strategies
- Report writing with AI
- Staying in control during EDA
- The minimalist version first

3/n

I've been using these techniques in my projects and with my teammates for the past year to compress experimental timelines from days to minutes. 2/n

Teaching an "Agentic Data Science" workshop to raise funds for SciPy 2026.

Compress weeks of work into a day using coding agents.

March 27 & April 3, 2-4 PM EST
Google Meet
40 spots, $1,000 min donation paid directly to NumFOCUS (tax-deductible)

Sign up: buff.ly/P1T2mh3 1/n

How to build self-improving coding agents - Part 3 written by Eric J. Ma on 2026-01-19 | tags: agents ai workflows productivity skills

Treat your coding agent like a teammate, not a chatbot. I think the trick is knowing when to update AGENTS.md vs create a skill. I share my simple model for evolving agent workflows and building real leverage. What’s your approach? Comment or share! buff.ly/FluyXjj #.aiagents #.devworkflow

How to build self-improving coding agents - Part 2 written by Eric J. Ma on 2026-01-18 | tags: agents ai skills mcp workflows

Tired of repeating the same coding steps? Skills—markdown playbooks—make agents smarter and your life easier. I share how skills saved me time and made team know-how reusable. Would you try skills? Comment or repost! buff.ly/rTPYdMf #.ai #.automation

Tired of repeating yourself to coding agents? Agents should actually learn from our feedback. AGENTS .md and simple playbooks helped my agents improve week by week—less babysitting, more progress. Tried this? Let me know or repost! buff.ly/5qi4pic #.ai #.automation

ICE murdered Renée Good in broad daylight. Less than 3 weeks later, they killed Alex Pretti, shooting him 10 times. Every day, we watch as people are ripped from their cars, their homes, their lives.

We can't allow ourselves to look away from this cruelty. Abolish ICE.

How I fixed a browser selection bug with sequence alignment algorithms written by Eric J. Ma on 2026-01-06 | tags: javascript bioinformatics katex canvas algorithms bugfix highlighting selection web development ui

Bioinformatics fixed my browser bug. DNA sequence alignment beat string normalization for messy text highlighting. Smith-Waterman, a DNA algorithm, solved my canvas-chat issue. Cross-domain ideas really work! Ever tried a wild fix? Comment or share! buff.ly/wnq3Ewz #.webdev #.algorithms

First contact with GLM-4.7 and... yeah, it feels a lot like Opus 4.5. One-shotting coding problems. I gotta probe where its limits are.

Same goes for product-minded engineers!

Canvas Chat: A Visual Interface for Thinking with LLMs written by Eric J. Ma on 2025-12-31 | tags: ai llm opencode claude visualization tools productivity

Ever wish LLM chats could branch like your thoughts? Linear chat holds us back—so I built Canvas Chat for nonlinear exploration. Open-source, infinite canvas for branching and merging ideas with Claude, GPT, and more. Repost or comment if this sparks ideas! buff.ly/KaFK4fR #.ai #.opensource

How I Themed My tmux with OpenCode + Claude (And When to Switch Models) written by Eric J. Ma on 2025-12-27 | tags: ai opencode claude tmux terminal creativity workflow pair-programming

AI themed my tmux bar—no plugins needed. AI let me focus on design, not syntax. Claude + OpenCode turned ideas into a terminal I love. You just need vision! How has AI changed your workflow? Repost or comment! buff.ly/aCXmjK1 #tmux #aipairprogramming

I use OpenCode, and have OpenCode read its own session as a way of remembering the exact history of what we did. I then feed AI the narrative I want to tell, and have it fill in the details :)

You Can Just Make Stuff with OpenCode and Claude Opus 4.5 written by Eric J. Ma on 2025-12-28 | tags: ai opencode claude automation workflow llm reasoning development review tools

I don't code anymore—I build. AI lets me ship projects by describing what I want, not how to do it. In 10 days, I shipped 6 projects with OpenCode + Claude Opus 4.5. The gap between idea and product is tiny now. Would you try this? Repost or comment! buff.ly/blroW24 #ai #buildfast

atggaacgccgctattgccatcgcattagcaccatggcgagcgcgaacgatcatgcgccgccgtataacgaatggtatgaagcgcgc!

(if you know, you know what to do!)

Two years of weekly blogging and what 2025 taught me written by Eric J. Ma on 2025-12-25 | tags: blogging retrospective coding agents llms bayesian biotech career writing marimo modal data science

🎄Merry Christmas to all! 104 blog posts. 2 years. Here’s what changed for me. Writing weekly turned AI into a true collaborator. Lessons on coding agents, Bayesian stats, and biotech leadership—shared honestly. What did 2025 teach you? Comment or repost! buff.ly/Kkebg57 #blogging #datascience

The selfish reason to do your best work written by Eric J. Ma on 2025-12-17 | tags: career growth work philosophy resilience reputation leadership mistakes success advice

Why do your best work if your company won’t notice? I think it’s about investing in yourself, not just your job. I share how building reputation and resilience pays off down the road. Ever had effort pay off later? Comment or share! buff.ly/SyXPuJc #careeradvice

Stop letting coding agents run wild—take control for better results. AI is a collaborator, not just a tool. My top tips: plan with AI, write docs/tests first, and use AGENTS.md as your agent’s guidebook. Got your own agent tips? Share below! buff.ly/kudW7pd #aiagents #productivity

3. A massive 24 year long Danish study demonstrated that, no, aluminium adjuvants in vaccines are not dangerous. They save around 4 million lives a year (11k lives A DAY.)

www.altmetric.com/de...

Explained in glorious detail by @sailorrooscout.bsky.social here: bsky.app/profile/sai...
4/8

What does it take to build a statistics agent? written by Eric J. Ma on 2025-12-02 | tags: statistics biotech reproducibility experiments research automation ai data open source bayesian

'Folk statistics' is costing science billions. AI can help, but only if we build it right. I built an AI stats agent for experiment design—sharing what worked, what failed, and why prompts matter. Faced these issues? Repost or comment! buff.ly/UGinNw9 #ai #statistics

How to Reference Code Across Repositories with Coding Agents written by Eric J. Ma on 2025-11-17 | tags: automation coding agents workflows productivity ai filesystem reference workspaces shell

Coding agents can access any repo—no workspace limits! Just give Copilot or Claude the file path and skip the hassle. I stopped copying files and juggling configs. It’s way simpler than I thought. Would you try this? Comment or share! buff.ly/EP7thlP #codingagents #productivity

How I Replaced 307 Lines of Agent Code with 4 Lines written by Eric J. Ma on 2025-11-16 | tags: llm graphs agents automation pocketflow llamabot python workflows abstractions state

307 lines of agent code—now just 4. Graph-based LLMs are clearer and more powerful. PocketFlow let me rebuild my AgentBot with a tiny, declarative graph. Cleaner code, instant visualization, and a new mindset for agents. Would you try this? Repost or comment! buff.ly/vQo50AX #llm #aiagents

That's amazing! I'm curious to hear how you store them -- is this in Notion, or Obsidian, or something else? And how do you enable their usage in situ, i.e. in context where you need them?

Safe ways to let your coding agent work autonomously written by Eric J. Ma on 2025-11-08 | tags: automation productivity coding agents safety workflow development prompting command line ai

Coding agents can wreck your repo in seconds. Auto-approve only read-only commands—never risky ones. I share my setup for safe agent autonomy and lessons learned. How do you keep agents in check? Comment or repost! buff.ly/PetnnVz #ai #automation

AI agents + Marimo notebooks = Python on autopilot. Letting AI write and check my Marimo code is a total workflow upgrade. The --watch flag and marimo check catch errors fast and keep me moving. Would you try this? Repost or comment! buff.ly/6LDN1RV #python #datascience