Christopher Gandrud

New short 📜 from me, Nicole Pashley, and @dom_val on THIS ONE WEIRD 90-YEAR OLD TRICK (Neyman allocation) that lets you leverage the pilot study of your experiment to maximize statistical efficiency. mattblackwell.org/files/papers/b…

Forget DALLE, PALM, M1 Pytorch. This is the biggest advancement in computer science this year: LaTeX math in github github.blog/2022-05-19-mat…

@conjugateprior @patrickdoupe @jon_m_rob Thanks for inviting me @conjugateprior. It was fun. I don't think I shared my slides: github.com/christophergan… . . .where I shared Fernández-Loría and Provost’s hot take. Probably the most thought provoking thing I’ve read recently.

I probably should have written this years ago, but here are some MLOps principles I think every ML platform (codebase, data management platform) should have: 1/n

What are things that work? Focusing on value generated for the business. Getting feedback from the team on what works & what doesn't. Eliminating biggest time wasters. Qualitative feedback. Investing in ways to create leverage for engineers. Iterating how the team/org works

Nice post by @jfkirn, one of the best PMs I've ever worked with, covering interesting aspects of experimentation at Lyft: - network effects - switching mindset from single decision to "always on" experiments - role of platform in aligning org decisions eng.lyft.com/challenges-in-…

New blog post! Discussing: 1. Whether a data scientist should be fullstack 2. What caused the unreasonable expectation that DS should know Kubernetes 3. An overview of the tools that can help abstract away infra to allow DS to own a project end-to-end 🚀 huyenchip.com/2021/09/13/dat…

To some extent, we all harbor this conceit that we alone could do something really amazing. But all the most impressive stuff is accomplished through some monumental coordination of talent and labor. Organizations are the best technology we've invented as humans.

Beta #pythoncode package for batch updating statistics (mean/variance). Useful for privacy preserving A/B tests: github.com/christophergan…

Notes on developing and distributing #python packages: elegant-heyrovsky-54a43f.netlify.app/workflow-for-p…

@ZalandoTech is looking for 2 senior backend engineers to develop and improve our high load A/B testing and continuous delivery system. Also, lots of opportunities to work with advanced data scientists to bring research to production. Get in touch at: grnh.se/1ee4e12b1us

We're looking for a leader who loves scaling experimentation, both with a high performance technical platform and through a product management organisation @ZalandoTech If you're interested, apply here: jobs.zalando.com/en/jobs/248746…

@ZalandoTech is looking for a leader who loves scaling experimentation, both in a high performance technical platform and through a product management organisation. If you're interested, apply here: grnh.se/4669712f1us

If I could build it again, I’d start with automating the evaluation of forecasts. It’s silly to build models if you’re not willing to commit to an evaluation procedure. I’d also probably remove most of the automation of the modeling. People should explicitly make these choices.

Slides from my Amazon Alexa lab fishbowl talk at bit.ly/2021ExPAlexaFi… Short 20 minute intro on #experimentation followed by Q&A on trends, culture, institutional memory, experts, OEC, platform, and mobile. #abtesting #hippo #experimentguide @kskripashankar

This work by @CMU_Stats PhD student @iwaudbysmith is *extremely* cool Surprisingly, you can do safe, anytime-valid causal inference - even in complex observational studies, where asymptotics usually dominate! Has important implications for modern online data collection & beyond

Does anyone else do this: I “like” content on twitter and even buy books on Amazon, not because I have a strong interest in that tweet or even reading that book, but I want to teach the recommendation algorithms that I like this type of stuff and want to see more.

If I had an hour to build a model, I’d spend 55 minutes building a fast and unambiguous evaluation procedure and 5 minutes trying out models.