Speaker
Francesca Lazzeri
Principal Group Director of Data and Applied AI Science @Microsoft, Advisor and Lecturer for MIT’s "Break Through Tech AI" Program
Full speaker profileAbstract
Modern enterprises want everyone, not just data engineers, to be able to ask questions of their data. But turning natural language into reliable, secure, and business-aware queries is far from simple. In this session, we introduce a new end-to-end architecture that combines intent screening, retrieval augmented context, hybrid fuzzy semantic matching, and large language model reasoning to generate executable queries that are both accurate and safe to run in production environments.
We will walk through how the system screens user intent, retrieves relevant past examples, aligns user language with schema entities, and constructs structured queries using a modular prompting strategy. We will also show how deterministic, LLM independent aggregation ensures numerical accuracy and eliminates hallucinated metrics. The talk will include a deep dive into the technical modules, vector search workflows, validation layers, and execution pipeline, along with performance and reliability considerations for real world deployments.
Attendees will leave with a clear understanding of how to design natural language to SQL systems that are secure, scalable, and enterprise ready, and how to combine LLMs with deterministic computation to deliver trustworthy insights at scale.
Main Takeaways
Natural language to SQL requires more than an LLM.Reliable systems must combine intent screening, contextual retrieval, and hybrid fuzzy semantic matching to correctly interpret user intent.
Deterministic computation is essential for trust.All aggregations and metric calculations must happen outside the LLM to eliminate hallucinations and ensure numerical accuracy.
Enterprise readiness depends on safety, validation, and governance. A production grade pipeline must enforce security checks, schema aware validation, and role based access controls before any query is executed.
Co-author of this session is Manpreet Singh, Senior Data & Applied Scientist, Microsoft.
Interview
Interview:
What is the focus of your work these days?
I am focused on building enterprise ready AI systems that combine large language models with deterministic, schema aware components. This includes designing architectures that safely translate natural language into executable queries, improving evaluation frameworks for AI assisted analytics, and ensuring that these systems meet the reliability, security, and governance standards required in large scale commercial environments.
What is the motivation behind your talk?
The motivation was to address a growing gap between what LLMs can generate and what enterprises can safely deploy. Many teams want natural language interfaces for data, but they struggle with accuracy, hallucinations, and security risks. This talk highlights a practical architecture that solves these challenges by blending LLM reasoning with retrieval, hybrid search, strict validation, and deterministic computation. The goal is to show a path to building systems that are both powerful and trustworthy.
Who is your session for?
This session is designed for data scientists, machine learning engineers, AI architects, and technical product leaders who are building or evaluating natural language interfaces for analytics. It is also relevant for engineers working on enterprise AI platforms, vector search, retrieval augmented generation, and evaluation frameworks for LLM based systems.