Eric J Ma's Website

In this blog post, I share why I believe the future of LLM applications lies beyond chat interfaces. Drawing on insights from colleagues, thought leaders, and my own experience building DeckBot, I argue that embedding AI into structured workflows—like TurboTax—creates more effective and delightful user experiences. Instead of relying on open-ended chat, we should inject LLMs at key moments within well-defined processes. Curious how this shift could transform the way we build and use AI-powered tools?

In this blog post, I share insights from my interviews with researchers and digital professionals on how to use AI, especially large language models, as a tool for active learning rather than passive consumption. I discuss strategies like creating personalized syllabi, applying critical thinking, and using AI for feedback, emphasizing that true learning requires effort and agency. Want to know the key trick to making AI your learning partner instead of your crutch?

In this blog post, I explore Cursor's latest update featuring automatic model routing, which eliminates the need for manual model selection. This change reduces the cognitive tax of micro-decisions, allowing me to focus more on coding. Drawing parallels to Apple's design philosophy, I discuss how removing unnecessary interfaces can enhance user experience. This shift in AI tool design suggests a future where systems handle complexity, simplifying user interactions. What other invisible frictions in AI tools could be automated away to improve our workflow?

In this blog post, I share how using Workspaces transformed my workflow while building out my tutorial repositories for SciPy 2025. By adding multiple repositories to a single workspace, I eliminated constant window switching and enabled my coding agent to access context across all repos simultaneously. This setup allows for seamless coordination between library code and tutorial examples. I provide a step-by-step guide to setting up a multi-repo workspace and offer tips for maximizing its benefits. Curious about how this can streamline your coding process and enhance your productivity?

In this blog post, I share my excitement for SciPy 2025 in Tacoma, Seattle. The Pacific Northwest's beauty, the vibrant community of tool builders, and the chance to give back through financial aid make it special. I'm thrilled to teach tutorials on LLMs and network analysis, and I can't wait for the code sprints, where collaboration and learning thrive. If you're considering attending, the opportunity to meet tool makers and explore new libraries is unmatched. Will you join us in this unique experience?

In this blog post, I share my experiences with Modal, a platform that simplifies cloud infrastructure. I discuss deploying web apps and hosting Ollama with low latency using Modal's cost-efficient, serverless solutions. I also explore creating a 'serverless workstation' with persistent storage and flexible resource allocation. Modal's intuitive volume management, unified abstractions, and easy debugging make it a standout choice for developers. Curious about how Modal can transform your cloud projects?

In this blog post, I share my experience from a live coding demo at BioIT World 2025, where I built a protein mass spectrometry calculator tool. I emphasize the importance of standardization in data science, starting with a design document, and using AI assistance for rapid development. Despite a live demo hiccup, I showcased the tool's capabilities and highlighted key lessons in AI collaboration, such as the value of context and interactive communication. How can AI tools enhance your development process while maintaining human oversight and creativity?

In this blog post, I share my experience with Marimo notebooks, highlighting their fully reactive nature and self-contained environments. I discuss how to run Marimo without installation using uv, and the benefits of AI-assisted coding. I also cover exporting notebooks to Markdown and deploying them as Modal apps. While Marimo's keybindings differ from Jupyter, its reactive execution and UI builder offer unique advantages. Curious about how Marimo can transform your coding workflow?

A high-throughput screening lab transforms their data analysis workflow by applying statistical thinking from the start. By combining robust estimation with R2D2 priors, they eliminate tedious manual data cleaning, automatically handle outliers, decompose sources of variation, and objectively measure both statistical estimation model and laboratory quality performance. This story demonstrates how thoughtful experimental design paired with principled statistical methods can dramatically improve both efficiency and scientific quality. How might statistical thinking transform your experimental workflow?

In this blog post, I explore how to tackle experimental noise and candidate ranking in protein screening using Bayesian methods. By employing R2D2 priors, we can decompose variance into interpretable components, helping us understand the true biological signal versus experimental artifacts. Additionally, Bayesian superiority calculation allows us to quantify the probability that one protein outperforms another, providing a more robust comparison than traditional methods. These techniques are not only applicable to protein screening but also to drug discovery, materials science, and more. Are you ready to enhance your experimental insights with Bayesian logic?