PhD Oral Exam - Saeed Fathollahzadeh, Computer Science

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Modern data systems demand substantial manual engineering across the data lifecycle: constructing dataset-specific ML pipelines and optimizing SQL queries for performance. Each stage resists automation due to dataset-aware reasoning requirements and the complexity of query transformations. This thesis introduces a suite of LLM-assisted frameworks unified by a common principle—generating candidate variants, either iteratively or as a set, then evaluating, verifying, and selecting among them. The central contribution is CatDB, a system for LLM-based, data-catalog-guided generation of data-centric ML pipelines. CatDB profiles datasets into a unified data catalog—capturing schema, feature types, value distributions, and statistics—and uses this metadata to construct dataset-specific rules guiding a zero-shot, in-context LLM prompt. An error management mechanism ensures generated pipelines are valid and correct. CatDB matches or exceeds state-of-the-art AutoML systems and LLM-based baselines at orders-of-magnitude lower pipeline generation latency. This thesis further contributes two SQL query optimization systems positioned on a spectrum from whole-query rewriting toward reusable, system-level rewrites. ReSequel generalizes queries into reusable templates, infers metadata-specific rewrite strategies, and verifies candidates for semantic equivalence—achieving speedups of up to 16× over native DBMSs and 22× over LLM-only baselines across eight benchmarks. ROLEX advances further along this spectrum, overcoming ReSequel's core limitation of online LLM dependency by repositioning the LLM as an offline knowledge generator—distilling workload rewrites into SMT-verified, optimizer-native rules with MCTS-guided ordering for deterministic, zero-LLM-cost rewriting at runtime. Together, these contributions establish a principled, LLM-augmented approach to automating the data-to-insight lifecycle while preserving correctness and competitive performance.