Shreya Shankar
PhD student in databases focusing on MLOps @Berkeley_EECS @UCBEPIC. Prev. 1st ML engineer @viaduct_ai, research @googlebrain, BS & MS @Stanford CS. She/they.
View on πThreads
While research papers certainly should split train/test temporally (their goals are to assess method validity), ML performance evaluation is a bit different for production (& shoul...
@bg_learning The most helpful thing for me has been a wide variety of datasets, models, applications in prod. But it has also been useful to have a basic understanding of linalg wh...
I don't think I've ever talked about this online, but the best ML-in-prod people I know have crazy good intuition for different classes of models can & can't do. We don't yet have...
Honestly: sometimes I feel defeated because ML observability is so hard. All facets are hard -- detecting, diagnosing, reacting to bugs. We don't have realtime ground truth labels...
studying for the Berkeley DB prelim after having built first-gen MLsys (feature stores, prediction serving, etc) on top of like postgres, Hadoop, Spark etc is kind of bonkers
i've hinted at this before -- @rogarcia_sanz and i are writing a paper on ML deployment practices. we're looking to interview more women-identifying practitioners who have deployed...
I did some DS & TL-ed DS projects. My answer: it's like what I imagine is an apprenticeship. Most classroom knowledge goes out the window, I traded import tf/pytorch for an intimat...
@rajiinio PN and I were having a nice long convo about this yesterday -- some CS research fields move so quickly (e.g., new communities spawned every few years, new "problems" inve...
I did elementary, middle, & high in public schools in Texas. What people don't know is that we have MANY shelter-in-place emergencies that don't make the news
One of my pet peeves with the public discourse on learning to code is the fixation on methodically reading books or taking courses. Most ppl I know who learned to code in K-12 did...
played around with postgresml dot org. in general i've drank the DB koolaid that ML will eventually be done in the RDBMS, but i still feel that the interfaces (BQ, too) are a bit c...
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