Projects & Impact

NL2Insights: Enterprise Text-to-SQL at Scale

2024 – 2026

Overview: NL2Insights is a fully automated pipeline for converting natural language to SQL queries, powering flagship IBM products and transforming how enterprises interact with their data.

Impact:

• 200,000+ SQL queries generated across 1,000+ databases
• Powers watsonx.data intelligence, BI Assistant, and Process Mining
• Doubled accuracy while reducing GPU usage
• Core component of IBM’s Data & AI strategy
• Production deployment across all IBM Cloud and AWS regions

Key Innovations:

• Optimized prompting strategies for enterprise data
• SQL safety modules ensuring query security
• Fine-tuned IBM Granite models for schema linking, content linking, and SQL generation
• Multilingual support (English, Japanese, with more languages coming)
• Automated reasoning capabilities for complex queries

Recognition:

• IBM Outstanding Technical Achievement Award - 2026 (NL2Insights Impacting Products and Clients)
• IBM Research Accomplishments (A-level) - 2025
• IBM Growth Award - 2025

BIRD Leaderboard: #1 in Text-to-SQL Benchmark

2023 – 2024

Achievement: Led IBM Granite Text-to-SQL models to first place in both tracks of the prestigious BIRD (BIg Bench for LaRge-scale Database Grounded Text-to-SQL Evaluation) leaderboard.

Challenge: BIRD features over 12,751 question-SQL pairs across 95 databases from 37 professional fields, emphasizing accuracy and execution efficiency.

Why It Matters:

• Outperformed GPT-4 and GPT-4o with smaller, more efficient models (Granite-20B and Granite-34B)
• Held #1 position for months against intense competition
• Sparked renewed interest from major players: Google, Alibaba, ByteDance
• Demonstrated IBM’s leadership in practical, production-ready LLM applications

Technical Approach:

• Fine-tuned models for three critical tasks:
  • Schema linking (connecting natural language to database schema)
  • Content linking (mapping to actual data values)
  • SQL code generation (producing accurate, executable queries)
• Developed novel training datasets from open-source and enterprise use cases
• Implemented advanced reasoning capabilities

Recognition:

• IBM Outstanding Technical Achievement Award - 2025
• IBM Research Accomplishments (A-level) - 2024
• Featured at THINK’24 (IBM’s flagship conference)
• Deployed on BAM and Watsonx.ai platforms

AutoDO: Automated Decision Optimization

2022 – 2023

Overview: An end-to-end automated system for solving sequential decision-making problems using data and knowledge-driven approaches.

Contributions:

• Designed application framework and system architecture
• Demonstrated effectiveness through comprehensive benchmarking
• Available on IBM API Hub portal for enterprise use

Recognition: Tutorial/Lab organizer at AAAI 2023: “Automated AI For Decision Optimization with Reinforcement Learning”

Evaluating Robustness in Multi-Agent Reinforcement Learning

2021 – 2023

Innovation: Proposed the first model-based adversarial attacks (cMBA) for cooperative multi-agent reinforcement learning.

Key Contributions:

• Novel victim agent selection strategy
• Constrained nonconvex optimization approach
• Comprehensive experiments on multi-agent MuJoCo environments

Publication: IEEE International Conference on Data Mining (ICDM) 2023

Patent: Filed patent application on systematic approach for evaluating robustness (Sep. 2022)

Federated Learning with Douglas-Rachford Splitting

2020 – 2021

Innovation: Proposed FedDR and asyncFedDR algorithms for federated learning with best-known communication complexity.

Key Features:

• Handles data heterogeneity across distributed clients
• Asynchronous updates for improved efficiency
• Rigorous theoretical guarantees

Publication: NeurIPS 2021 (35th Conference on Neural Information Processing Systems)

Impact: Advanced the state-of-the-art in federated optimization for non-convex problems