Self-Verifying AI Systems
When specifications are incomplete, AI systems may make incorrect assumptions and leave users to discover the failures after the fact.
Explore the projectBarzanMozafari
How can AI systems verify and improve their own work?
Self-Verifying AI and Self-Improving Data Systems
I build systems that verify and improve their own work—from AI applications to data platforms and cloud infrastructure.
Professor · AI Systems Researcher · Entrepreneur
Current Research
My current research explores how AI and data systems can verify, improve, and optimize their own work.
Self-Improving Data Systems
AI, database learning, and workload feedback for improving queries and data pipelines while preserving semantics and operational constraints.
GenRewrite
Uses large language models and counterexample-guided validation to rewrite SQL queries while preserving semantics.
DAGSmith
Uses pipeline dependencies as optimization signals and separates rewrite generation from equivalence checking.
Cloud Optimization & FinOps
Workload-aware methods for improving performance, resource efficiency, and cost across cloud data warehouses and infrastructure.
Keebo
Autonomously optimizes cloud data-warehouse performance and cost in production.
Writing
Technical perspectives on reliable AI, data systems, and optimization.
Why Slower SQL Can Make Your dbt Pipeline Faster
Modern analytics is increasingly organized as recurring SQL pipelines rather than isolated SQL statements. Tools like dbt have gained extreme popularity in recent years, allowing teams to write each transformation as SQL and make dependencies between transformations explicit. This has led to directed acyclic graphs (DAGs) with hundreds or thousands of interdependent SQL models. Traditional query optimizers and source-to-source query rewriters are not effective at optimizing these expensive pipelines. In this post, we explore how DAGSmith treats these explicit dependencies as optimization signals and uses an LLM to find pipeline-level refactorings and optimize a dbt project holistically.
Read article →News & Highlights
GenRewrite appears at SIGMOD 2026
GenRewrite introduces LLM-assisted query rewriting with iterative semantic validation
Automated warehouse optimization research presented at SIGMOD
Press & Media
Barzan Mozafari named Morris Wellman Faculty Development Professor
TIBCO introduces ComputeDB following its acquisition of SnappyData
Contention-aware transaction scheduling arrives in InnoDB
Selected Earlier Work
Efficient and Scalable Machine Learning
Methods for making machine learning more efficient, scalable, and robust under communication, computation, and memory constraints.
Database Learning: QuickSel and BlinkML
Database-learning methods that improve estimation and reduce computation using workload history.
Predictable Database Systems
Methods for diagnosing, bounding, and reducing performance variability under changing workloads.
Interactive Analytics at Scale: BlinkDB, VerdictDB, and SnappyData
Systems for interactive analytics over massive data with explicit guarantees on accuracy and response time.
Research Impact
Cloud-warehouse optimization commercialized through Keebo
Keebo developed the first autonomous cloud data-warehouse optimization platform and was used by Fortune 100 companies.
BlinkDB ideas commercialized through SnappyData and TIBCO
BlinkDB was commercialized through SnappyData, which was later acquired by TIBCO.
Transaction schedulers adopted by MySQL and MariaDB
Our CATS algorithm has become the default transaction scheduler in MySQL and Percona Server for MySQL, estimated to be running on hundreds of millions of deployments worldwide. Our earlier VATS algorithm was also adopted as the default transaction scheduler in older versions of MariaDB.