AetherDialect: Towards a Deterministic Text-to-SQL

The School of EECS is hosting the following DS Seminar:

AetherDialect: Towards a Deterministic Text-to-SQL

Lessons from building an open-source text-to-SQL engine — and the open research problems we'd like to work on together.

Speakers: Dr. Roozbeh Derakhshan, Akul Ameya and Charlotte Birkinshaw 

Chair: Dr Zijian Wang

Abstract: Text-to-SQL, translating natural language questions into executable queries over relational data, is one of the most actively researched problems in NLP and data systems today, and one of the most aggressively pursued in industry. New benchmark results, model architectures, and commercial products appear almost monthly. Yet despite the headline numbers, production deployments in enterprise settings remain unreliable. LLM-based systems hallucinate joins, fabricate columns, drift across schema versions, and fail silently on semantically subtle queries. In domains where correctness is non-negotiable — finance, healthcare, manufacturing, government — "usually right" is not an acceptable foundation for a decision-support system.

In this seminar we present AetherDialect, an open-source Text-to-SQL engine developed at Data & Knowledge Enterprise (DKE) in Brisbane. We will briefly survey the dominant academic and industry approaches, schema-linking models, in-context learning, agentic and self-correcting LLM pipelines, and explain why each remains fundamentally probabilistic. AetherDialect takes a different stance: rather than generating SQL and verifying it after the fact, we compile the question into a structured analytical intent and validate it against the schema, controlled joins, and curated business definitions before any SQL is produced. Ambiguity is surfaced explicitly. The same question on the same data yields the same answer. The LLM operates inside a deterministic envelope, not as a free generator of executable code.

The talk will outline the architecture and implementation of AetherDialect, walk through how determinism is preserved end-to-end, and close with a set of open research problems where we are actively seeking collaboration with the university — including formal guarantees on intent compilation, evaluation methodology beyond execution accuracy, and federated query across sovereign data boundaries. AetherDialect is now released as open source; attendees are welcome to engage with the codebase.

Speakers:

Dr. Roozbeh Derakhshan is the Founder and Director of Data & Knowledge Enterprise (DKE), a Brisbane-based AI and data company. He launched DKE in 2021 to make data and AI accessible to organisations of all sizes, particularly those constrained by limited IT capacity and budget. Roozbeh has worked on stream and real-time data processing from its early academic origins through to production infrastructure, and still writes code and ships product personally. His career spans IBM, SAP, and Accenture in industry, and ETH Zürich and the University of Queensland in research.

Akul Ameya is a Valedictorian Data Science graduate (CGPA 4.0) and leads the development of AetherDialect's deterministic Text-to-SQL library. He works across machine learning, data systems, and LLM engineering, and has several accepted and ongoing papers targeting venues including NeurIPS and ACL. He approaches problems with a systems-first mindset — determinism, validation, and scalability, grounded in strong mathematical foundations.

Charlotte Birkinshaw is a Software/AI Engineer at DKE, working on data pipelines (Azure Databricks), cloud platform engineering (Azure, AWS), and end-to-end delivery of scalable data platforms. Her honours research in Prof. Pauline Pounds' lab at UQ produced a web-based data entry portal for a sepsis PCR diagnostic, integrated with Azure Databricks. She is currently extending that engagement into a natural-language querying layer over the diagnostic dataset.