Constrained generation undergirds most LLM workflows, but most folks don't take the time properly ensure instructions are followed. Reya Vir collected a great dataset on inferring constraints from custom prompts and using them as verifiers to steer generations.

🔍How do we make sense of messy, real-world documents? We (the EPIC Lab at UC Berkeley) are building DocETL, an open-source system for LLM-native data processing. We've started to create a showcase of demos. Here's our first: an analysis of public feedback on US AI strategy. 🇺🇸

Shreya Shreya Shankar is one of the world's leading experts on LLM evals.. she's able to speak fluently to both academics and practitioners alike - and has led a bunch of foundational research in this area (spade, evalgen, promptevals). Check this out!

do you want to use LLMs to process your data for ~10x less while **guaranteeing** the quality of gpt-4o or an expensive LLM? check out our work on BARGAIN!

Low-cost LLM-powered data processing at scale **with guarantees** - now open-source! Get accuracies that are nearly that of your SOTA model at a fraction of the 💸. Take a look and let us know what you think.

The EPIC data lab out of Berkeley just released an open source tool called BARGAIN. We can use small models to do the work of an expensive LLM if we can predict when the small model is right. This is the magic behind Downlink. Here's the demo:

✨ Grateful and excited to share what’s ahead. I’ll be joining University of Utah as an Assistant Professor starting Fall 2026, as part of the UtahDB Lab! I'll continue working data summaries and the integration of LLMs and DBs. Before that, I’ll be spending a year back at Snowflake❄️.

1/3 The US didn’t end up leading the world in computing by luck. It happened because it made long-term, public investments in basic research, especially through NSF. That’s what created the technology that today’s companies are built on.

I came to the US on an F-1 student visa. Since then, I’ve built a company worth over $20 billion—headquartered in Pennsylvania. Student visas are one of America’s greatest talent funnels. Let’s not close the door behind us.

Excited to partner with the PromptQL folks on this important benchmarking effort around data agents!

In this work, we show how to use off-the-shelf LLMs to generate code for accelerators. This is interesting as accelerators are often "low resources," i.e., there isn't much code written using such accelerators to train custom models. Check out our paper for details!

🚨BREAKING: New Python library for agentic data processing and ETL with AI Introducing DocETL. Here's what you need to know:

A decade of HILDA! Still a vibrant sub community of databases, human in the loop is more relevant than ever From Arnab’s keynote.

.United Airlines is it your official policy to discriminate against parents of young children? Even though we boarded early in UA 59, (about to take off) in group 1 (thanks to status) and had found space overhead for our stroller, one of your staff came later and forced us to check it in

DocETL transforms mountains of unreadable business documents into actionable insights using AI-powered semantic analysis 📚 process unstructured text data without building task-specific models or manual annotation 🎧 thedataexchange.media/docetl/