DSPy's abstractions worked for us, but its runtime didn't
DSPy is a useful AI workflow development tool, but its runtime configuration introduced issues for our production workflows.
acompa blog
Notes and updates from acompa
Latest posts
What seems to be happening after Claude Opus 4.6
How the emergence of useful agentic coding in late 2025 might affect everything.
On job-hunting in 2024 as a machine learning engineer
Lessons from 39 job conversations after a layoff from Spotify, and what the market looks like for ML engineers in 2024.
Large-scale data-driven initiatives at Airbnb
How Airbnb scaled its data infrastructure with Airflow, Zipline, and other frameworks for ML at hundreds-of-models scale.
A survey of ACM RecSys 2018
Trends from RecSys 2018: contextual bandits, deep learning for recommendations, and fairness in personalization.
A review of SciPy 2017, including my talk on music trends
Speaking at SciPy 2017 about modeling trends in music, plus highlights from other talks at the conference.
Git for data science at iHeartRadio
Teaching git fundamentals to iHeartRadio's algorithms team to scale collaboration beyond single-owner folders.
Introductory notes on regularization
A primer on L1, L2, and ElasticNet regularization and why they help models generalize.
Beyond Trending Topics: identifying important conversations in communities
Detecting unusual spikes in Twitter community activity using time series decomposition, built at Betaworks with Rohit Jain.
Presenting leptoid at Surge 2012
Presenting an autoscaling project at Surge 2012 and learning that good instrumentation matters more than a good forecasting model.
Installing R and rpy2 from scratch
A step-by-step guide to installing R and rpy2 on a bare EC2 instance for Python developers who need R's statistical packages.
Sleeplessness => bad code
How sleep deprivation during a busy summer led to silly bugs in a toy virtual memory manager.
Using Item Response Theory Instead of Traditional Grading to Assess Student Proficiency
Why traditional grading ignores question difficulty, and how item response theory offers a better measure of student ability.