90% Text2SQL accuracy from 52%
107 live use cases including supply gap analysis & executive summaries
Standardized Text2SQL pipeline integrated into an open-source AI library
Reduced report generation time from one week to just hours for faster critical decision-making
Summary
The conversational data agent for supply chain and procurement that delivers domain-specific insights, smart recommendations, and scalable, user-friendly analytics represents a transformative shift from simple data retrieval to an agentic transformation of a Silicon Valley Big Tech leader’s management of its complex global logistics.
By implementing an agentic Text2SQL pipeline with a sophisticated ecosystem of autonomous agents, the company and Stratpoint have built a system that bridges the gap between non-tech supply chain planners and enterprise-level data, transforming a historically slow, manual reporting process into a near real-time self-service tool with an improved Text2SQL accuracy of 90% from 52%.
Challenges
The company’s supply chain planners are masters of resource allocation and logistics, but they are typically non-technical users who do not write code or interface directly with SQL databases to pull real-time information from SAP. Historically, extracting insights from SAP required manual data preparation and long wait times, often resulting in a week-long delay for critical reports that dictate billions in potential profit or loss. This encouraged the company to create a tool that allows a planner to query in natural language and receive a data-backed answer instantly.
In its initial state, the conversational data agent’s prototype was confined to a limited AWS EC2 environment that could only support 30-50 users with in-memory data tables. To achieve enterprise-wide adoption, this prototype required a significant scalability upgrade. The objective extended far beyond basic data access; the goal was to build a highly transparent, multi-agent translation system that supply chain teams could trust with high-stakes decisions.
The conversational data agent aimed to solve this scalability gap by creating a text-to-mass-text interface, allowing natural language queries to be converted directly into SQL statements for real-time analysis. The company sought Stratpoint’s expertise in the implementation and deployment of the AI agent solution, with consideration for performance, accuracy, and security.
Solutions
To perform the heavy lifting and high-stakes experiments required to move the project from a local proof-of-concept to a production-grade application, the company worked with Stratpoint’s specialized AI team to augment their internal engineers. The team consisted of a program manager, an AI solutions architect, a DevOps solution architect, a senior AI engineer, three AI engineers, and a DevOps engineer.
While the company’s engineers provided direction, Stratpoint’s AI team focused on building the evaluation pipelines, prompt engineering, and the infrastructure necessary for an on-premise rollout.
Architectural Modernization
A critical phase of this vision was the transition of the conversational data agent’s infrastructure to meet enterprise-level demand. The project transitioned from a local proof of concept to a production-grade application deployed within the company’s on-premise data cluster. This move was critical for maintaining data security and addressing the high-performance needs of the planning department.
- AI factory migration: The team successfully moved the architecture to the company’s on-premise data cluster, utilizing high-performance, bleeding-edge computer hardware to reduce latency and deliver total data sovereignty.
- Microservice strategy: To support massive, parallel development and future-ready scalability, the team transitioned the system from a monolithic structure to a microservices architecture, separating the front-end, back-end, and MCP server.
- Massive data connectivity: The solution integrates with SAP HANA and Databricks ODP via SQL endpoints, allowing users to query massive datasets like the Waterfall Data Feed and Demand Lead Time (DDLT) tables.
- DevOps maturity: The team established formal CI/CD pipelines and containerized services to support parallel development and faster deployment cycles. The team built an automated, secure software delivery process that covers validation, testing, security scans, CVE remediation, deployment, monitoring, and incident response.
- End-to-end evaluation: Stabilized the end-to-end evaluation pipeline to ensure that, as new use cases were added, accuracy remained high.
The Agentic Engine
The true breakthrough is a sophisticated multi-agent ecosystem where multiple specialized agents work in tandem to manage complex data retrieval and compensation workflows. While the complete pipeline is extensive, it is driven by these core foundational agents:
Acts as the translator by interpreting complex user intent, rephrasing ambiguous queries, and clarifying what the user is actually looking for.
The trust builder that dismantles fear of black-box AI by providing a step-by-step reasoning process that explains how every answer is derived, building confidence among users.
A vital, self-healing quality-control layer that validates SQL output, automatically detects execution errors, and regenerates the query with error-context feedback to ensure accuracy.
This explainability allows planners to see exactly which filters and logic were applied to their data, transforming the conversational data agent into a transparent data retrieval assistant.
Smart Recommendations and Visual Analytics
The agent functions as a conversational AI interface where planners can query supply chain data and receive insights in text, graph, or chart form.
- Detailed vs. directional responses: UI upgrades that allow users to choose the depth of the AI’s analysis.
- Supply gap analysis: Specific filters were added to support transactional and inventory tables across different planning departments and supply chain scenarios, enabling planners to visualize where inventory shortages might occur.
- Sequential query workflows: The tool supports multi-step problem-solving workflows, moving away from simple one-off queries to complex, sequential data-fetching logic.
Technologies Used
