Eric J Ma's Website

Encountering a PyPI upload error related to license metadata? The solution is straightforward - switch from setuptools to Hatchling as your build backend. In this post, I walk through how to fix the "license-file introduced in metadata version 2.4" error by updating your pyproject.toml configuration. Along the way, I learned some new things, including the fact that modern build backends like Hatchling provide better support for PEP 621 metadata features compared to older tools like setuptools.

When building machine learning models in biology, we often encounter data that's been heavily processed with statistical transformations like p-values and normalizations. This essay argues that this practice fundamentally undermines our ability to build reliable models and maintain interpretable datasets. Through a real-world example of protein binding experiments, it demonstrates why collecting physical quantities (like binding strength in nM) with proper replicates is vastly superior to statistical artifacts, and how Bayesian estimation can help us properly handle experimental variation while maintaining physical units. Are you tired of wrestling with hard-to-interpret biological data and ready to build more reliable experimental pipelines?

In this blog post, I explore how modern Python tooling is flipping the script on the age-old "which Python should I use?" question. Through my experience with uvx, marimo, and juv, I show how we're moving away from the traditional headache of environment setup and toward a world where tools automatically ship you the exact Python you need. No more environment setup puzzles – just specify your Python version and get straight to work. It's a liberating shift that's changing how I approach one-off Python work, and I think it's pretty exciting!

In this blog post, I share my journey of tackling dependency bloat in LlamaBot. What began as a simple LLM bot framework had grown into a monolithic system with an extensive dependency chain, leading to massive installation sizes. By mapping dependencies, refactoring the code, and organizing optional dependencies, I managed to reduce the container size significantly. This exercise taught me the importance of regular codebase maintenance and focusing on core functionalities. Now, LlamaBot is leaner and more efficient. Curious about the strategies I used to achieve this transformation?

In this blog post, I explore the concept of 'what makes an agent' by discussing various implementations of LlamaBot, a Python package for LLM exploration. I dissect the differences between SimpleBots, StructuredBots, and AgentBots, highlighting their capabilities and limitations in terms of agency and decision-making. Through audience discussions and examples, I aimed to provoke thought on the definition and design of agents, and together, we had an engaging discussion. Can we truly define an agent, or is it like the Turing Test, a concept that evolves with our understanding and technological advancements?

In this blog post, I explore the challenges biotech teams face when integrating public datasets with internal data for machine learning. Despite initial excitement, issues like data compatibility, missing variables, domain shifts, and biological complexity often arise. I suggest a shift from a machine learning perspective to a decision support approach, advocating for separate models and a decision fusion layer that incorporates human expertise. This method respects the complexity of biological systems and aids in effective decision-making. How can we better navigate these challenges to accelerate biotech discoveries?

When typing became physically demanding, I discovered that dictation tools could do more than just help me write – they could fundamentally change how I capture and develop ideas. Using Better Dictation and VoicePal, combined with AI assistance, I found a way to write that matches the natural flow of thought. This isn't just about accessibility or working around limitations; it's about finding a better way to translate the nonlinear, rapid-fire nature of our thoughts into written words. I share my approach to preserving authentic voice while using AI tools, and why sometimes constraints push us toward unexpected improvements in how we work.

In this post, I share a practical approach to managing secrets in data science workflows, learned from personal experience with both successes and mistakes. I cover essential tools like direnv and .env files for local development, strategies for secure secret handling in Jupyter notebooks, and crucial version control practices including pre-commit hooks to catch accidental API key commits. I also discuss team collaboration approaches for secret sharing, platform-specific secrets management features, and what to do when secrets accidentally get committed to repositories. While tools like AWS Secrets Manager exist for enterprise needs, I focus on practical, accessible methods that create robust security through layered defenses, following proven software development principles that apply equally well to data science work.

In this blog post, I explore what defines an LLM agent, highlighting its goal-oriented non-determinism, decision-making flow control, and natural language interfaces. I also discuss when to use agents, emphasizing the importance of variable scope inputs and constrained actions. By examining industry perspectives from Anthropic and Google, I also explore how agents can effectively handle diverse inputs while maintaining defined action boundaries. Real-world examples, like a bill calculation bot and a literature research assistant, illustrate these principles. How can these insights transform your approach to designing agent applications?

In this blog post, I reflect on my year-long challenge of writing a blog post every week, surpassing my goal with 53 posts. This journey taught me the power of consistency, improved my ability to communicate complex ideas, and helped me develop AI-assisted tools to streamline my workflow. I also explored the intersection of life sciences and computation, aiming to accelerate scientific discovery. How did these experiences shape my approach to integrating AI into creative processes and what insights can you gain from my journey?