Published research
Co-authored a study on postpartum health
Published in JMIR Formative Research.
Blog
Practical thinking on AI strategy, ML product development, and building AI features that actually work.
Recent
QA isn't coming back, it's splitting
AI can handle the QA a separate team used to do. What it can't handle is domain judgement, and that belongs closer to the product.
Read more →Stop making up eval examples
Why eval suites built on invented examples teach you almost nothing, and what to do instead.
Read more →The problem with "garbage in, garbage out"
Why the most repeated line about data quality usually leads to premature optimization.
Read more →Sharing my Maintainable Software Podcast episode!
I joined Robby Russell to talk about what makes maintaining AI and ML systems different from traditional software.
Read more →Sebastian Raschka is a gift
Why I make time to work through a whole technical book with a reading group.
Read more →Co-authored a study on postpartum health
Published in JMIR Formative Research.
Read more →Get some clarify about your role!!
A simple exercise for leaders who feel responsible for everything: write down what you own, then go align on it with your founder.
Read more →Use LLMs to build features, classical ML to predict
A pattern that works well for predictive models: let LLMs extract signal from messy unstructured data, then let a model like XGBoost make the actual prediction.
Read more →Why I love graphs
My lifelong obsession with nodes and edges.
Read more →The discomfort of writing software right now is a green flag
Everyone is fumbling with new AI-powered tooling. If your most senior engineers are comfortable, that's the thing to worry about.
Read more →Two themes from AI Dev
My takeaways from the AI Dev Conference in San Francisco
Read more →How to evaluate technical papers for your company
How to cut through the volume of AI research: don't try to reason your way to an answer, design a small experiment that tells you whether a paper matters for your product.
Read more →Passionate ambivalence: hosting Sue Heilbronner on Never Ask for the Sale
My product book club and a few founders crowded into my living room for a talk with Sue Heilbronner about selling without the hard sell, limiting beliefs, and distilling your essence into 30 seconds.
Read more →The classic product playbook doesn't work for AI
Why PM-defines-problem then engineering-builds-it breaks down on genAI and ML projects, and the iteration rhythm that replaces it.
Read more →Notes on Extreme Multi-label Classification
What I took away from the PRIME paper on classifying into enormous label spaces, and why its momentum-updated label centroids are the interesting part.
Read more →Notes on Claude
What I've been seeing, what I've changed my mind about, and what I'm recommending, from worktrees to Cowork to Claude's Constitution.
Read more →I used to be horrendous at giving feedback
I let go of every one of my early hires because I couldn't deliver the feedback they needed. Here's the approach I use now.
Read more →Red flags when building AI
Common red flags that lead to AI builds dragging on indefinitely or features falling flat on launch, and how to distinguish them from the expected discomfort of building probabilistic systems.
Read more →A case for CTOs writing code
Why CTOs should carve out time to write code despite conventional wisdom, and how to do it without becoming a bottleneck.
Read more →Golden Datasets: The Essential First Step for AI-Powered Apps
Why creating a golden dataset of curated user questions paired with verified correct answers is a critical first step before building AI-powered applications.
Read more →Jobs to be done for ML features
Why the placement of an ML feature in the user journey matters more than model choice, using the Jobs-to-Be-Done framework to define what your model should optimize for.
Read more →UX Patterns for AI Features
Why chat interfaces aren't the future of AI UX, and how AI patterns like smart defaults, contextual suggestions, and proactive helpers create better experiences.
Read more →Data shapes design
One of the biggest mindset shifts when building AI features: moving from design-first to data-first, where your data determines what's feasible, stable, and valuable.
Read more →Working with me
How I help startup founders and tech leads bring clarity and momentum to their AI efforts, from figuring out what's worth building to shipping features users trust.
Read more →You seriously need to look at the data
Why hand-labeling data and manually inspecting errors are some of the highest-ROI activities when building AI features, despite teams resisting the manual work.
Read more →ML metrics in production
Why standard uptime monitoring isn't enough for AI-powered features, and the minimum metrics teams should track to catch model drift and degradation.
Read more →Minimum data needs for ML models
How much data do you need to build a predictive model? The honest answer is you don't know upfront, but starting with 100-200 labeled examples is a solid approach.
Read more →Customer discovery for AI features
Before building an AI feature, make sure someone will actually use it. Key takeaways from The Mom Test for better customer discovery.
Read more →How to use LLMs for Classification
A practical approach to using LLMs as classifiers by combining multi-criteria LLM judgments with a lightweight model like XGBoost for better accuracy and calibration.
Read more →Your architecture decision is probably wrong, and that's okay
Why there's rarely a single 'best' technical decision, and how decision records, speed, and iterative review help engineering organizations get better over time.
Read more →Common failure modes of machine learning projects
Five common failure modes that cause ML projects to fail, from mismatched expectations and lack of progress to inadequate data, overly-complex solutions, and lack of expertise.
Read more →The case for 1-week sprints for ML teams
Why 1-week sprints are uniquely powerful for ML and research teams, despite seeming counterintuitive for open-ended experimental work.
Read more →