Panasonic Avionics Links GPT to SQL Server for Automated Forecasting

Harnessing SQL Server and AI for Analytics With Panasonic Avionics Corporation — Ai4

Ai4

Restricting LLMs to database schemas while running predictive models locally allows non-technical staff to pull analytics instantly without leaking enterprise data.

AI-assisted summary of Harnessing SQL Server and AI for Analytics With Panasonic Avionics Corporation — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 29, 2025, this presentation features Vijay Panwar, Senior Software Engineer at Panasonic Avionics Corporation. According to the page description, the session examines the transformative potential of AI-driven chatbots and OpenAI's GPT models to enhance Microsoft SQL Server interactions, enabling users to pose natural language inquiries and receive optimized queries alongside real-time insights.

## TL;DR Vijay Panwar demonstrates an enterprise solution that bridges non-technical business users and Microsoft SQL Server using a lightweight Python Streamlit chatbot. The system translates plain-English prompts into optimized SQL queries, subqueries, and index recommendations by sharing only database schemas with LLMs to preserve strict data privacy. Furthermore, queries involving timestamped data automatically trigger a local CatBoost machine learning model to generate 10-to-12-month predictive forecasts with 95% accuracy, eliminating engineering bottlenecks for managerial reporting.

## Atmosphere & style The presentation is pragmatic, technical, and approachable, marked by engaging live software demonstrations and lighthearted banter with the audience over the difficulty of writing SQL subqueries. Panwar speaks with confident authority and accessible enthusiasm, focusing directly on real-world business hurdles and data security boundaries. It feels like an encouraging, problem-solving tech briefing that emphasizes democratizing data access without sacrificing enterprise governance.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=B7vAz3AGt34&t=4s) Integrating LLMs with Microsoft SQL Server allows non-technical business stakeholders to query enterprise data using plain English without manual SQL writing. - [01:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=112s) The system architecture ensures strict enterprise data privacy by passing only structural database schemas to OpenAI, keeping all row-level data strictly local. - [01:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=112s) A lightweight Streamlit frontend in Python eliminates the need for dedicated UI development teams while enabling one-click query generation and data downloads. - [01:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=112s) An automated machine learning pipeline powered by CatBoost triggers whenever time-series columns are detected, generating 10 to 12-month forward-looking forecasts with ~95% accuracy. - [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) The AI assistant automatically constructs complex nested subqueries and joins that are typically challenging for non-engineers to write manually. - [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) Enterprise guardrails enforce read-only execution to prevent destructive write, update, or delete commands from executing against the database. - [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) The platform proactively recommends missing database indexes and provides ready-to-paste DDL scripts to create views and stored procedures for heavy tables. - [06:51](https://www.youtube.com/watch?v=B7vAz3AGt34&t=411s) LLMs generate SQL syntax exclusively, while predictive forecasting is performed locally by dedicated machine learning algorithms rather than the language model itself. - [08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) The architecture remains LLM-agnostic, enabling organizations to switch between OpenAI GPT, Microsoft Copilot, Meta LLaMA, or DeepSeek based on compliance and infrastructure needs. - [08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) An autonomous conversational AI agent is planned to deliver natural language performance summaries directly instead of returning raw data tables. - [08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) Self-service natural language database querying removes engineering bottlenecks, enabling managers and directors to retrieve targeted reports immediately.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=B7vAz3AGt34&t=4s) Introduction and Speaker Background

Vijay Panwar, Senior Software Engineer at Panasonic Avionics Corporation, opens the session by welcoming attendees and introducing the core topic: harnessing Microsoft SQL Server with AI chatbots to empower users who lack SQL expertise. He outlines his professional credentials, highlighting nearly 12 years of industry experience, 12 published research papers centered on AI and Web API system optimization, a senior IT membership with Sigma Xi, and winning an MIT hackathon the prior year.

The presentation addresses the universal challenge of organizational data access: while every enterprise relies heavily on data, many business stakeholders lack the technical proficiency to write SQL queries. By integrating OpenAI’s GPT models with SQL Server, non-technical users can pose natural language questions in plain English—such as asking for performance records over the last two years—to automatically generate and run optimized SQL queries and download results with a single click.

- Speaker holds ~12 years of industry experience and 12 published research papers. - Full member of Sigma Xi and past winner of an MIT hackathon. - Emphasizes that business users can retrieve complex data without writing manual queries.

```mermaid flowchart LR accTitle: High-Level Natural Language to SQL Architecture subgraph InputStage["User Interface"] A["User Query in Plain English"] --> B["Streamlit Chatbot UI"] end subgraph AIStage["AI & Query Generation"] B --> C["OpenAI GPT Model (Schema Only)"] C --> D["Optimized SQL Query"] end subgraph DataStage["Execution & Analysis"] D --> E["SQL Server Execution"] E --> F["Real-Time Insights & ML Forecast"] end ```

---

[01:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=112s) System Architecture, Security, and ML Forecasting

Panwar contrasts traditional SQL workflows—plagued by manual, time-consuming query construction for non-developers—with an AI-enhanced natural language processing pipeline. Crucially, the architecture strictly protects sensitive enterprise data: zero row-level records or customer values are shared with OpenAI; only the structural database schema is transmitted so the model understands table relationships and data types.

The user interface is built using Streamlit in Python, eliminating the need for dedicated UI development teams. When queries retrieve time-series or timestamp-based columns, an integrated machine learning pipeline automatically triggers to generate predictive forecasts covering the next 10 to 12 months.

```mermaid flowchart TD accTitle: Detailed AI and ML Processing Flow subgraph UI["Streamlit Frontend"] U["User Enters Question"] end subgraph SchemaLayer["Security Boundary"] S["Database Schema (No Row Data)"] --> O["OpenAI GPT Model"] end subgraph QueryEngine["Query & Execution"] U --> O O --> Q["Generate SQL + Index Recommendations"] Q --> DB[("SQL Server")] end subgraph PostProcessing["Analytics Engine"] DB --> CHK{"Timestamp Column Detected?"} CHK -- "Yes" --> ML["CatBoost ML Model (95% Accuracy)"] CHK -- "No" --> RES["Direct Result Display"] ML --> RES end ```

- **Data Privacy**: Only database schema metadata is shared with OpenAI; data rows remain strictly local and secure. - **Frontend**: Lightweight Python Streamlit interface. - **Predictive ML**: Automated CatBoost time-series forecasting triggered by timestamp columns.

---

[03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) Live Demonstration: Optimization, Subqueries, and Security Guardrails

The presentation demonstrates the chatbot UI in action. The left navigation pane displays all connected database tables. When asked natural language prompts like "give me the top 10 high-value customers", the system generates appropriate SQL joins and immediately recommends missing indexes to optimize query execution.

Panwar engages the audience by asking if writing subqueries is simple or difficult; an audience member answers that it is simple, to which Panwar humorously responds, "you must be an engineer then." He explains that writing subqueries—such as finding the second-highest employee salary or retrieving the top 10 records while excluding a specific group—is typically challenging for non-engineers, but the chatbot handles exclusions and nested logic automatically. Furthermore, the system provides enterprise safeguards and database recommendations: - **Read-Only Safety**: Prevents destructive operations; requests to update, insert, or delete records are blocked. - **Index Suggestions**: Suggests performance indexing (e.g., expanding from 2 to 7 index suggestions as query complexity increases). - **View Recommendations**: Suggests creating database views and stored procedures instead of querying heavy tables directly, generating the exact view creation scripts ready to paste into SQL Server.

---

[06:51](https://www.youtube.com/watch?v=B7vAz3AGt34&t=411s) Real-World Use Cases and CatBoost Predictive Analytics

Panwar details enterprise use cases across banking and finance (real-time customer segmentation) and retail analytics (instant sales insights spanning five-year historical records). When time-stamped records are queried, the system invokes an embedded CatBoost gradient boosting model that achieves roughly 95% predictive accuracy.

Addressing an audience question on whether ChatGPT handles the forecasting directly, Panwar clarifies: "No, no, this is not ChatGPT. This is an LLM [pipeline]." ChatGPT is solely responsible for generating the SQL syntax, while local Python code inspects the result set; if a date or timestamp column is present, the CatBoost ML model runs locally against the retrieved data to project trends for the subsequent 10 to 12 months.

| Aspect | Query Generation (LLM) | Predictive Forecasting (ML) | |:--- |:--- |:--- | | **Model** | OpenAI GPT (or alternative LLMs) | CatBoost Regressor / Forecaster | | **Input** | Plain-English prompt + DB schema | Query result set with timestamps | | **Output** | Executable SQL query & indexing tips | 10–12 month forward-looking predictions | | **Accuracy** | 100% syntactical execution parity | ~95% statistical forecasting accuracy | | **Data Scope** | Structural schema only | Local SQL Server data records |

---

[08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) Implementation Best Practices and Upcoming Autonomous Agent

Panwar outlines best practices for deploying SQL-AI systems in enterprise environments: - **LLM Flexibility**: While OpenAI is used in this deployment, teams can alternate with Microsoft Copilot, Meta LLaMA, or DeepSeek depending on architectural and compliance requirements. - **Security & Compliance**: Zero personal identifiable information (PII) or business records are sent externally. - **Validation**: Generated queries were tested directly against SQL Server, confirming 100% execution parity.

Looking ahead, Panwar announces work on an autonomous AI agent for SQL Server scheduled for release by the end of the year. This conversational agent allows natural, human-like dialogue—enabling users to ask qualitative questions like "how is Vijay Panwar performing over the last two years?" and receive plain-language summaries backed by automated backend analytics rather than raw CSV tables.

He concludes by noting that the core chatbot model presented will be published open-source for free by the end of the month, emphasizing the productivity gains: "If a manager wants to see some report, they don't need to bug the engineer. They can just go and ask the question and the report is there immediately and you can download it." Senior executives and directors can independently pull targeted reports without waiting for engineering backlogs.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=B7vAz3AGt34&t=4s) Harnessing SQL Server and AI for Predictive Analytics: Practical Methods and Real-World Use Cases — Vijay Panwar

> [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) give me the top 10 high-value customers — Vijay Panwar

> [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) you must be an engineer then. — Vijay Panwar

> [06:51](https://www.youtube.com/watch?v=B7vAz3AGt34&t=411s) No, no, this is not ChatGPT. This is an LLM [pipeline]. — Vijay Panwar

> [08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) how is Vijay Panwar performing over the last two years? — Vijay Panwar

> [08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) If a manager wants to see some report, they don't need to bug the engineer. They can just go and ask the question and the report is there immediately and you can download it. — Vijay Panwar

## Mentioned Vijay Panwar, Panasonic Avionics Corporation, Microsoft SQL Server, OpenAI, GPT, Streamlit, Python, CatBoost, Sigma Xi, MIT, Microsoft Copilot, Meta LLaMA, DeepSeek, Ai4

## Video description Vijay Panwar, Senior Software Engineer, Panasonic Avionics Corporation

Abstract

In my presentation titled "Harnessing SQL Server and AI for Predictive Analytics: Practical Methods and Real-World Use Cases," I will examine the transformative potential of AI-driven chatbots in enhancing SQL Server interactions through natural language processing, and facilitating effortless database queries. This discussion will demonstrate the integration of OpenAI’s GPT models with SQL Server, empowering users to pose intricate database inquiries in straightforward English and receive optimized SQL queries along with real-time insights.

[Source](https://www.youtube.com/watch?v=B7vAz3AGt34)

中文简报

松下航空电子将GPT连接至SQL Server实现自动化预测 将大语言模型限制在数据库架构层面并在本地运行预测模型，使非技术人员无需泄露企业数据即可即时提取分析结果。 Vijay Panwar展示了一套企业级解决方案，通过轻量级Python Streamlit聊天机器人搭建起非技术业务用户与Microsoft SQL Server之间的桥梁。为严格保护数据隐私，该系统仅向大语言模型共享数据库架构，即可将日常英语提示转换为经过优化的SQL查询、子查询及索引建议。此外，涉及时间戳数据的查询会自动触发本地CatBoost机器学习模型，生成准确率达95%的10至12个月前瞻性预测，消除了管理报表制作中的工程瓶颈。 Panasonic Avionics 高级软件工程师 Vijay Panwar 展示了将 OpenAI GPT 模型与 Microsoft SQL Server 结合的方案，让非技术业务人员能够直接使用日常自然语言查询企业数据库并获取洞察。 (0:04) 系统架构严格保障企业数据隐私，仅向 OpenAI 传输数据库 Schema 结构元数据，实际业务行级数据完全保留在本地进行计算与查询。 (1:52) 系统具备只读安全防护机制以拦截修改与删除操作，可自动构建复杂的嵌套子查询与表联结，并主动提供索引优化与视图创建脚本建议。 (3:52) 大语言模型仅负责生成 SQL 语法，若查询结果包含时间戳列，本地集成的 CatBoost 机器学习模型将自动运行，提供准确率约 95% 的未来 10 至 12 个月趋势预测。 (6:51) 该核心聊天机器人模型计划于当月免费开源，团队还计划在年底前推出面向 SQL Server 的自主 AI Agent，支持更深度的人性化业务问答与报告生成。 (8:53)

Reading summary

Vijay Panwar, Senior Software Engineer at Panasonic Avionics Corporation, outlines an architecture that pairs large language models with Microsoft SQL Server to let non-technical stakeholders query enterprise databases in plain English without exposing sensitive row-level data. Schema Isolation and Privacy Safeguards (1:52) Traditional enterprise workflows often stall because business users lack the SQL expertise required to extract data independently, burdening engineering teams. Panwar's architecture resolves this by using a Python Streamlit interface that forwards user prompts to OpenAI's GPT models alongside structural database schema metadata rather than actual table contents. Under this design, zero row-level records, customer metrics, or personally identifiable information leave the enterprise perimeter. The external language model parses table relationships and column data types purely to write syntactically correct queries, which then run internally against SQL Server. Automated Query Generation and Execution Guardrails (3:52) The system translates plain-language requests—such as identifying top-tier customers or extracting records with specific exclusions—into functional SQL joins and subqueries without manual developer input. Panwar observes that while nested subqueries and exclusions present steep technical hurdles for non-engineers, the model automates the logic behind operations like salary rankings or segmented filtering. To safeguard enterprise databases, the pipeline implements built-in security guardrails that restrict interactions to read-only queries, automatically blocking destructive update, insert, or delete actions. The system also acts as an administrative advisor: it suggests index additions to boost execution performance and outputs ready-to-run scripts for creating views or stored procedures to keep heavy queries off production tables. Decoupled Predictive Analytics via CatBoost (6:51) Panwar clarifies that the generative AI component does not conduct forecasting; its role ends once valid SQL syntax is produced. When SQL Server returns results, local Python code inspects the dataset for date or timestamp attributes. If a timestamp column is detected, the pipeline automatically routes the returned records to a local CatBoost gradient boosting model. This local model produces 10- to 12-month forward projections at approximately 95% statistical accuracy, allowing use cases such as retail trend projections or financial customer segmentation to run securely without sending raw historical figures to an external vendor. Model Flexibility and Autonomous Agent Roadmap (8:53) Panwar notes that enterprise deployments are not strictly tied to OpenAI; organizations can swap in models such as Microsoft Copilot, Meta LLaMA, or DeepSeek to accommodate institutional compliance standards. In internal testing, generated queries achieved 100% syntactical execution parity on SQL Server. Panasonic Avionics planned to release the core chatbot model as free open-source software by the end of the month, with an autonomous conversational agent scheduled for release by year's end. The planned agent aims to support dialogue-based qualitative summaries—such as multi-year performance assessments—enabling directors and managers to obtain synthesized answers directly instead of waiting on engineering backlogs for raw CSV extracts. Panwar illustrates how separating external schema parsing from internal ML execution can democratize data access while maintaining strict data governance. However, the operational success of such systems remains contingent on consistent schema interpretation across complex schemas and the reliability of automated forecasting when applied arbitrarily to dynamic database outputs.

精炼阅读摘要

Panasonic Avionics Corporation 高级软件工程师 Vijay Panwar 提出了一种将大语言模型与 Microsoft SQL Server 深度集成的分析架构，旨在帮助非技术人员通过自然语言完成复杂数据检索与分析。该方案通过将数据库架构元数据与底层行级数据严格物理隔离以保障合规，并结合本地机器学习算法实现自动化时间序列预测。 架构安全边界与元数据隔离机制 (1:52) Panwar 指出，传统企业数据调取严重依赖工程人员手写 SQL，导致业务部门面临较高的沟通与时间成本。为此，该方案采用基于 Python 的 Streamlit 搭建轻量级交互前端，后端接入大语言模型处理自然语言提问。在底层安全设计上，该架构设定了明确的数据隔离红线：系统仅向大语言模型传输数据库架构（Schema）元数据，绝不向外部模型上传任何具体的行级业务记录、客户数据或个人身份信息（PII）。 这种元数据隔离设计确保了敏感业务资产完全保留在企业本地环境中。在模型选型上，除演示所用的 OpenAI GPT 系列外，系统亦支持接入 Microsoft Copilot、Meta LLaMA 或 DeepSeek 等替代模型，企业可根据数据合规与基础设施要求灵活替换。此机制在消除公有云模型数据泄露隐患的同时，保留了模型解析表关联与数据类型的能力。 自然语言转SQL与数据库优化防护 (3:52) 针对业务人员难以独立编写多表连接、复杂排除逻辑或多层嵌套子查询（如筛选特定员工薪资层级）的问题，该系统实现了端到端的自然语言转换。用户输入常规英语指令（例如查询前十名高价值客户），系统即可自动生成带有正确连接与过滤条件的 SQL 语句，并在测试中验证了与 SQL Server 的完全语法执行一致性。 除了生成查询，系统还内嵌了企业级安全防护与数据库性能调优建议机制。在权限与安全性方面，系统设置了严格的只读防护屏障，自动拦截并阻止 INSERT、UPDATE、DELETE 等任何破坏性写入操作。在性能优化方面，系统能够根据查询复杂度动态推荐缺失的索引配置（例如根据复杂度由 2 个扩展至 7 个索引建议），并能主动建议创建视图（View）或存储过程以规避全表扫描，甚至可直接生成可供执行的建表与视图脚本。 分层分析与CatBoost预测引擎 (6:51) 针对现场关于预测机制的疑问，Panwar 明确澄清大语言模型仅负责生成 SQL 语法，并不直接参与任何数据推理或数值预测。实际的预测处理交由本地数据流水线执行：本地 Python 代码检测查询结果中是否存在日期或时间戳字段，一旦识别到时间序列数据，便会自动触发内置的 CatBoost 梯度提升机器学习模型。 据演讲者介绍，该本地 CatBoost 模型基于 SQL Server 检索出的真实数据运行，专门用于生成未来 10 至 12 个月的前瞻性趋势预测，且统计预测准确率据称可达约 95%。这种双层分析流水线将 LLM 的自然语言交互能力与成熟机器学习算法的结构化预测能力解耦，支持诸如零售行业五年历史销售洞察以及银行业务中的实时客户细分等场景。 业务自助赋能与自主代理演进路线 (8:53) 该系统的核心商业价值在于破除企业内部跨部门取数的效率瓶颈。业务经理与总监级管理人员无需向工程师提交工单排队，便可通过日常语言提问实时调取指标并一键下载报告，从而压缩管理决策链路并释放技术人员的开发精力。 在开源与产品演进规划方面，Panwar 表示该系统的核心 Chatbot 方案将在当月底免费开源。此外，其团队正在研发一款面向 SQL Server 的自主 AI 代理（Autonomous Agent），计划于年底前发布。该代理旨在支持更具定性维度的自然语言多轮对话（例如直接询问某员工近两年的综合表现），并由系统自动完成后台指标聚合，直接输出结合业务语境的文字摘要而非单纯的原始表格。 该方案展示了将 LLM 严格限制在元数据理解与语法生成边界、将数据计算与时序预测留由本地引擎处理的实用架构思路。然而，其落地成效仍高度依赖企业既有数据库元数据的规范性，且演讲中宣称的 95% 预测准确率与零行级泄露保障，在复杂异构的企业生产环境中仍需经过持续的泛化与边界安全验证。

Detailed notes — zh

## Background 该视频由频道 Ai4 于 2025 年 8 月 29 日在 YouTube 平台发布。内容为松下航空电子（Panasonic Avionics Corporation）高级软件工程师 Vijay Panwar 的主题演讲，主要探讨如何将 OpenAI 的 GPT 模型与 SQL Server 深度整合，通过自然语言对话实现低门槛的数据库查询、索引优化与实时预测分析。 ## TL;DR 演讲系统介绍了利用 AI 对话机器人赋能非技术业务人员直接与 SQL Server 交互的端到端解决方案。架构在 Streamlit 前端与本地数据库之间运作，仅将 Schema 元数据暴露给大模型以确保企业核心数据 100% 留存本地，并具备严格的只读安全防护与智能索引调优机制。当下游返回数据包含时间序列时，系统会自动无缝触发 CatBoost 机器学习模型进行未来 10 至 12 个月的高精度趋势预测。Panwar 最后分享了实施该方案的四大最佳实践，并公布了开源计划与新一代自主 AI Agent 的研发路线。 ## Atmosphere & style 整场演讲呈现出严谨务实与生动风趣并存的现场氛围，讲师凭借深厚的技术积淀，将复杂的 SQL 嵌套子查询与机器学习流水线拆解得清晰易懂。过程中穿插着与现场技术同行的幽默互动以及对传统报表管理流程的善意调侃，既有企业级落地架构的技术硬度，又充满了解放业务生产力的实用主义激情。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=B7vAz3AGt34&t=4s) 通过将大语言模型与企业本地 SQL Server 深度整合，非技术业务人员能够直接使用自然语言完成复杂的数据库查询与报表提取。 - [00:54](https://www.youtube.com/watch?v=B7vAz3AGt34&t=54s) 架构坚守“零私有数据上云”的安全原则，仅向大语言模型传输表结构 Schema，所有数据行均保留在本地执行，规避数据泄露风险。 - [00:54](https://www.youtube.com/watch?v=B7vAz3AGt34&t=54s) 前端基于 Streamlit 快速搭建轻量化交互界面，用户无需专业前端支持即可自定义查询粒度并实现结果的一键 CSV 导出。 - [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) 系统内置严格的安全白名单机制，自动拦截并彻底阻断包括 DELETE、UPDATE 和 INSERT 在内的破坏性写入操作，保证数据库只读安全。 - [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) AI 助手可自动推断跨表 JOIN 关联并精准构建复杂的嵌套子查询，大幅降低非工程师群体处理高阶 SQL 逻辑的门槛。 - [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) 具备实时的数据库性能调优能力，系统能根据查询计划动态推荐缺失的关键索引，并自动生成视图（View）与存储过程代码模板。 - [06:51](https://www.youtube.com/watch?v=B7vAz3AGt34&t=411s) 后端集成时间序列自动拦截器，一旦识别出查询结果包含日期字段，即刻自动拉起 CatBoost 机器学习链路生成未来 10 至 12 个月的趋势预测。 - [06:51](https://www.youtube.com/watch?v=B7vAz3AGt34&t=411s) 大语言模型仅负责根据语义生成 SQL 语句而不接触或直接计算数据，后续的时序预测完全依托本地 CatBoost 算法在真实数据上独立运行。 - [08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) 底层模型具备高扩展性，架构除 OpenAI 外还兼容 Microsoft Copilot、开源 Llama 以及 DeepSeek 等多种大语言模型。 - [08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) 未来路线将向 SQL Server 原生 AI Agent 演进，通过全拟人化自然语言直接口头汇报综合业务洞察，免去用户阅读复杂报表和数据的负担。 - [11:26](https://www.youtube.com/watch?v=B7vAz3AGt34&t=686s) 打破跨部门提需求的效率壁垒，业务人员与管理者无需依赖工程团队排期写脚本，即可实现随时随地的自主数据探索与决策。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=B7vAz3AGt34&t=4s) 讲师背景与议题引入：利用 AI 解锁 SQL Server 来自松下航空电子（Panasonic Avionics Corporation）的高级软件工程师 Vijay Panwar 向现场观众致谢并开宗明义地指出了本次演讲的核心主旨：如何驾驭 SQL Server，并通过智能对话机器人（Chatbot）赋能那些完全没有 SQL 基础的业务人员。在正式展开技术探讨前，Panwar 简要介绍了自己的专业履历。他拥有近 12 年的工业级研发经验，在人工智能（AI）与 Web API 系统性能优化等领域发表了约 12 篇学术研究论文；他不仅是 IEEE 高级会员，也是著名科学研究荣誉学会 Sigma Xi 的正式成员，并在去年斩获了麻省理工学院（MIT）黑客松竞赛的冠军。 紧接着，Panwar 勾勒了今天分享的整体脉络，内容将涵盖 SQL Server 与 AI 的融合机制、数据库 Schema 识别、查询优化等关键技术环节。他指出，数据是当今每家企业的生命线，各业务部门都渴望获取并分析数据，然而现实中大量业务人员根本不知道该如何编写复杂的 SQL 查询语句。借助将 AI Chatbot 与 SQL Server 深度整合的方案，用户只需使用最朴素的日常英语提问——例如“能给我看一下 Vijay Panwar 过去两年的表现数据吗？”——系统便能自动调用 OpenAI 生成对应的 SQL 语句，并在企业私有服务器上执行该查询，最终用户只需轻点一下鼠标即可导出所需数据。 - 讲师背景：12 年工业界软件工程经验，发表 12 篇关于 AI 与 Web API 的论文，IEEE 高级会员，Sigma Xi 荣誉学会成员，MIT 黑客松获奖者。 - 核心愿景：打破技术壁垒，让完全不懂 SQL 语法的非技术人员能够通过自然语言与大型数据库顺畅交互。 ### [00:54](https://www.youtube.com/watch?v=B7vAz3AGt34&t=54s) 传统查询痛点与端到端架构工作流 传统数据库交互面临着极其显著的痛点：编写手动 SQL 查询不仅门槛高，而且对于非技术用户来说极度耗时。虽然技术团队能够自如地处理脚本，但跨部门的提需求与沟通往往造成严重的效率瓶颈。而引入基于自然语言处理（NLP）的智能架构后，不仅能实现实时的 SQL 生成，还能提供即时的查询优化建议（如索引推荐），从而大幅提升决策效率；更重要的是，用户可以随心所欲地控制数据粒度，例如直接指定“只返回我关心的这三列，而不是整个宽表”。 为了彻底打消企业在数据安全方面的顾虑，Panwar 着重强调了该方案的隐私合规设计。许多企业担心商业机密泄露，而在该架构中，“哪怕是数据库里的单行数值，也绝不会发送给 OpenAI；一切数据在本地都享有 100% 的安全性”。系统唯一与外部大模型共享的只有数据库的结构定义（Schema），因为如果 AI 不清楚表结构和字段定义，就无法精准编写出语法和逻辑正确的 SQL 语句。整个前端交互采用 Streamlit 构建，纯 Python 代码即可快速渲染出简洁直观的用户界面，完全不需要额外投入专业前端开发人员；当下游查询结果包含时间序列字段时，系统还会自动拉起机器学习预测链（LLM Chain / ML Pipeline），对未来 12 个月的趋势进行前瞻推演。 ```mermaid flowchart TD accTitle: 自然语言到 SQL Server 的安全查询与预测工作流 subgraph Client ["用户交互层 (Streamlit UI)"] A["用户输入自然语言问题"] F["结果展示与一键 CSV 下载"] end subgraph CloudAI ["大语言模型服务 (OpenAI)"] B["接收数据库 Schema<br/>(绝不传输任何数据行)"] C["理解意图并生成<br/>优化后的 SQL 查询"] end subgraph InternalDB ["企业本地环境 (SQL Server)"] D["执行 SQL 查询并评估索引"] E{"结果集是否包含<br/>时间戳字段？"} G["CatBoost 预测模型<br/>(输出未来 10-12 个月预测)"] end A --> B B --> C C --> D D --> E E -- "是" --> G E -- "否" --> F G --> F ``` ### [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) 智能查询生成、安全防护与实时索引优化 在展示实际原型时，用户界面的左侧侧边栏直观罗列了当前已连接的所有数据表。用户只需在对话框中键入自然语言指令，系统便会即刻生成对应的 SQL 语句、拉取数据表，并同步给出性能调优建议。Panwar 在演示中特别指出系统的防注入与权限隔离：“如果有人在对话框里试图输入‘帮我删除这个’、‘帮我更新那个’或者‘插入一条记录’，Chatbot 是绝不会执行的，系统在权限层面做到了百分之百的只读与安全防范。” 针对更复杂的场景，例如用户提问“找出消费金额最高的前 10 名高净值客户”，系统不仅能够自动构建涉及多表关联（JOIN）的查询语句，还会自动进行静态与动态执行分析，提示当前数据库缺少哪些关键索引以避免全表扫描。Panwar 此时转向现场观众进行互动调侃：“你们平时写子查询（Subquery）感觉吃力还是简单？只需要回答简单还是难就行。”当听到台下一位观众轻描淡写地回答“简单”时，Panwar 幽默地回应道：“那你一定是个软件工程师了！” 随后，他进一步向非工程师群体阐释了子查询的复杂性：它本质上是在一个查询之上再套一层查询。例如为了找出公司里薪资“第二高”的员工，传统写法必须依靠嵌套子查询来排除第一名或者使用窗口函数及 LIMIT 限制；而借助这个 AI 助手，用户只需自然地说出“请给我前十名，但我希望在名单中排除某些特定人群”，AI 就能迅速写出精准的子查询嵌套逻辑。系统不仅会依据数据规模动态提升索引建议（例如由简单查询的 2 个索引建议增加至 7 个索引建议），甚至还会主动提议：“你是否需要为此类高频数据生成一个数据库视图（View）？”随后自动生成创建视图的代码，工程师只需复制并在 SQL Server 中执行即可，极大地减轻了底层运维与建模负担。 - 安全与权限策略：内置严格的 SQL 操作白名单机制，完全阻断 `DELETE`、`UPDATE`、`INSERT` 等破坏性写操作。 - 自动化 SQL 建议：自动推断跨表 JOIN 关联，根据查询计划从推荐 2 个索引动态扩展至 7 个关键缺失索引，并提供一键生成视图（View）和存储过程（Stored Procedure）的代码模板。 ### [06:51](https://www.youtube.com/watch?v=B7vAz3AGt34&t=411s) 结合 CatBoost 的时间序列预测与落地场景 谈及商业落地价值，Panwar 列举了金融银行业的实时客户细分（Customer Segmentation）以及零售行业针对历史销售表现的即时洞察。业务人员可以快速调取过去 5 年的完整历史数据并一键导出。在此基础上，Panwar 的架构整合了 CatBoost 机器学习预测算法，在模型评测中实现了高达 90.95%（约 95%）的高预测精度，能够为企业业务未来走势提供极其稳健的指引。 现场有观众对预测机制提出了疑问，以为这是由 ChatGPT 直接凭空生成的预测数值。Panwar 立即进行了明确的技术澄清：“不，这并不是 ChatGPT 算出来的。ChatGPT 的职责仅仅是根据自然语言生成 SQL 查询语句，它根本看不到也不获取实际数据。”真正的运作逻辑在于后端的执行检测拦截器：当 SQL 引擎执行完毕并回传数据后，系统代码会自动扫描结果集的所有列属性；一旦识别出数据包含时间戳或日期字段，后端的机器学习预测链路便会瞬间无缝介入。Panwar 在幻灯片中展示了基于 2012 年至 2013 年的历史数据库所生成的未来 10 到 12 个月（即 2013 年至 2014 年）走势图。预测曲线与真实数据曲线几乎高度重合，这证明了企业完全可以依托这套系统对客户增长、销售额或内部 KPI 进行高可信度的中远期推演。 - 核心预测算法：采用针对表格数据极其高效的 CatBoost 梯度提升模型。 - 预测触发机制：代码层自动化侦测返回列的时间戳属性，零人工干预触发预测流水线。 - 准确率表现：在给定的历史客户与销售验证集上，未来 10 至 12 个月的时序预测准确度达到 95%。 ### [08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) 最佳实践、未来路线图与业务赋能 在系统落地实施与推广的最佳实践方面，Panwar 总结了四大核心支柱： 1. **数据库准备**：规范 SQL Server 的元数据命名与 Schema 结构，为大模型意图解析打好地基； 2. **灵活选型底层大模型**：尽管他个人偏好使用 OpenAI，但该架构完全兼容 Microsoft Copilot、开源的 Llama 以及 DeepSeek 等前沿模型，只要针对具体业务场景做好输出交叉校验即可； 3. **数据隐私与合规性**：坚守“零私有数据上云”原则，仅暴露元数据 Schema，全流程满足企业严苛的安全审查； 4. **持续监控与准确性迭代**：对 AI 生成的所有 SQL 语句在真实的 SQL Server 环境中反复测试与性能基准比对，经检验，模型在测试用例中的结果集准确率达到了 100%。 关于下一步的产品演进，Panwar 透露他正在全力研发更先进的“SQL Server 原生 AI Agent（智能体）”。当前的 Chatbot 虽然强大，但输出的形式仍然偏向表格和数据指标；而新一代的智能体将具备更深层次的人机拟人化交互能力，“即便是小孩子也能毫无障碍地向它提问”。面对那些对 CSV 文件或冰冷数据表格完全不感兴趣的管理层，新智能体可以直接在后台对全局数据进行深入扫描与综合归纳，直接用流畅的自然语言口头汇报“Vijay Panwar 在过去两年的整体绩效与业务达成情况”。Panwar 宣布，本次展示的 Chatbot 源码已计划在当月底面向全社会完全免费开源，而功能更加全面的 AI Agent 完整架构预计将在今年年底前正式对外发布。 最后，Panwar 总结了该方案为企业生产力带来的颠覆性改变。它将沉重的数据库交互真正变成了“触手可及的指尖体验”。他微笑着向现场的管理层打趣道：“如果今天在座的有各位经理，无意冒犯，但现实中如果经理需要紧急调阅某份报表，真的不需要再去麻烦工程团队加班写脚本了。你们只需打开系统打字提问，报表瞬间呈现，点击下载按钮即可完成交付。”无论是基层业务分析师还是高级总监，都能够借助这项技术实现数据获取与分析的真正自主。 根据视频片段及背景信息，由于提供的转录文本片段（第 11:26 至 11:29）仅包含短短几字且属于视频尾声的残缺语句，以下为基于该片段内容的忠实记录与梳理： ### [11:26](https://www.youtube.com/watch?v=B7vAz3AGt34&t=686s) 数据库自主查询与权限限制 在演讲的结尾部分，Vijay Panwar 补充说明了 AI 驱动的自然语言数据库查询系统在实际工作流中的灵活运用。通过将 OpenAI 的 GPT 模型与 SQL Server 进行深度整合，业务人员和开发者能够摆脱对传统层级审批或管理人员命令的依赖。当团队成员产生超越经理日常业务要求的个性化数据分析需求时，他们不再需要专门向管理层提出开报表申请，而是可以“直接前往系统自行执行查询”（"just go and do"），快速获取实时优化的 SQL 语句与数据洞察。 - **自主数据探索**：员工可针对非例行、非经理强制要求的数据需求进行自主提问与查询。 - **效率提升**：借助自然语言处理（NLP），降低了从提出数据问题到生成 SQL 语句的门槛，减少了中间沟通成本。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=B7vAz3AGt34&t=4s) 能给我看一下 Vijay Panwar 过去两年的表现数据吗？ — Vijay Panwar > [00:54](https://www.youtube.com/watch?v=B7vAz3AGt34&t=54s) 哪怕是数据库里的单行数值，也绝不会发送给 OpenAI；一切数据在本地都享有 100% 的安全性 — Vijay Panwar > [00:54](https://www.youtube.com/watch?v=B7vAz3AGt34&t=54s) 只返回我关心的这三列，而不是整个宽表 — Vijay Panwar > [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) 如果有人在对话框里试图输入‘帮我删除这个’、‘帮我更新那个’或者‘插入一条记录’，Chatbot 是绝不会执行的，系统在权限层面做到了百分之百的只读与安全防范。 — Vijay Panwar > [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) 你们平时写子查询（Subquery）感觉吃力还是简单？只需要回答简单还是难就行。 — Vijay Panwar > [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) 那你一定是个软件工程师了！ — Vijay Panwar > [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) 请给我前十名，但我希望在名单中排除某些特定人群 — Vijay Panwar > [03:52](https://www.youtube.com/watch?v=B7vAz3AGt34&t=232s) 你是否需要为此类高频数据生成一个数据库视图（View）？ — Vijay Panwar > [06:51](https://www.youtube.com/watch?v=B7vAz3AGt34&t=411s) 不，这并不是 ChatGPT 算出来的。ChatGPT 的职责仅仅是根据自然语言生成 SQL 查询语句，它根本看不到也不获取实际数据。 — Vijay Panwar > [08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) 即便是小孩子也能毫无障碍地向它提问 — Vijay Panwar > [08:53](https://www.youtube.com/watch?v=B7vAz3AGt34&t=533s) 如果今天在座的有各位经理，无意冒犯，但现实中如果经理需要紧急调阅某份报表，真的不需要再去麻烦工程团队加班写脚本了。你们只需打开系统打字提问，报表瞬间呈现，点击下载按钮即可完成交付。 — Vijay Panwar > [11:26](https://www.youtube.com/watch?v=B7vAz3AGt34&t=686s) just go and do — Vijay Panwar ## Mentioned Vijay Panwar, Panasonic Avionics Corporation, SQL Server, OpenAI, GPT, Streamlit, CatBoost, Microsoft Copilot, Llama, DeepSeek, IEEE, Sigma Xi, 麻省理工学院 (MIT) ## Video description Vijay Panwar, Senior Software Engineer, Panasonic Avionics Corporation Abstract In my presentation titled "Harnessing SQL Server and AI for Predictive Analytics: Practical Methods and Real-World Use Cases," I will examine the transformative potential of AI-driven chatbots in enhancing SQL Server interactions through natural language processing, and facilitating effortless database queries. This discussion will demonstrate the integration of OpenAI’s GPT models with SQL Server, empowering users to pose intricate database inquiries in straightforward English and receive optimized SQL queries along with real-time insights. [Source](https://www.youtube.com/watch?v=B7vAz3AGt34)

Source quotes

[0:04] Vijay Panwar: 能给我看一下 Vijay Panwar 过去两年的表现数据吗？ [0:54] Vijay Panwar: 哪怕是数据库里的单行数值，也绝不会发送给 OpenAI；一切数据在本地都享有 100% 的安全性 [0:54] Vijay Panwar: 只返回我关心的这三列，而不是整个宽表 [3:52] Vijay Panwar: 如果有人在对话框里试图输入‘帮我删除这个’、‘帮我更新那个’或者‘插入一条记录’，Chatbot 是绝不会执行的，系统在权限层面做到了百分之百的只读与安全防范。 [3:52] Vijay Panwar: 你们平时写子查询（Subquery）感觉吃力还是简单？只需要回答简单还是难就行。 [3:52] Vijay Panwar: 那你一定是个软件工程师了！ [3:52] Vijay Panwar: 请给我前十名，但我希望在名单中排除某些特定人群 [3:52] Vijay Panwar: 你是否需要为此类高频数据生成一个数据库视图（View）？ [6:51] Vijay Panwar: 不，这并不是 ChatGPT 算出来的。ChatGPT 的职责仅仅是根据自然语言生成 SQL 查询语句，它根本看不到也不获取实际数据。 [8:53] Vijay Panwar: 即便是小孩子也能毫无障碍地向它提问 [8:53] Vijay Panwar: 如果今天在座的有各位经理，无意冒犯，但现实中如果经理需要紧急调阅某份报表，真的不需要再去麻烦工程团队加班写脚本了。你们只需打开系统打字提问，报表瞬间呈现，点击下载按钮即可完成交付。 [11:26] Vijay Panwar: just go and do

Source quotes

[0:04] Vijay Panwar: Harnessing SQL Server and AI for Predictive Analytics: Practical Methods and Real-World Use Cases [3:52] Vijay Panwar: give me the top 10 high-value customers [3:52] Vijay Panwar: you must be an engineer then. [6:51] Vijay Panwar: No, no, this is not ChatGPT. This is an LLM [pipeline]. [8:53] Vijay Panwar: how is Vijay Panwar performing over the last two years? [8:53] Vijay Panwar: If a manager wants to see some report, they don't need to bug the engineer. They can just go and ask the question and the report is there immediately and you can download it.

Original transcript

[0:04] Good afternoon everyone and thank you so [0:06] much for joining me here. [0:09] Uh I will be talking about harnessing [0:11] the SQL server and how a chatbot can [0:16] help anyone who don't have SQL server [0:19] knowledge. [0:21] Okay. So before we start uh that's me. I [0:26] have almost 12 years of industry level [0:28] experience right now. I published almost [0:31] 12 research paper which is focusing on [0:33] AI and web API how we can leveraging and [0:36] can improve the system and we can go [0:38] from there. Senior IT member full member [0:41] of Sigma 11 and I just won the hackathon [0:45] from MIT uh last week not last week last [0:48] year actually. Thank you. [0:54] Okay. So today agenda I will be trying [0:56] to introduce the SQL server and AI [0:59] integration. There will be a technical [1:02] approach like schema recognization, [1:04] query optimization and everything. I [1:07] know like the data is like key part of [1:10] each and everyone. Every company want to [1:12] gather the data. There will be a people [1:14] who had no idea how to write a SQL [1:16] queries. But by having this chatbot [1:19] integrated with your SQL server, you can [1:21] ask a simple English question like can [1:24] you give me a data for VJ punar? How he [1:27] did in last two years? The open AAI will [1:30] generate a SQL query for you. It ran the [1:33] SQL query against your server, generate [1:35] the data and on one click you can able [1:37] to download it and I will share how the [1:40] schema works after this slide. But [1:43] nothing or not even a single value from [1:46] your data is going to share with OpenAI. [1:48] Everything is completely secured. [1:52] So there is a couple of challenges in [1:55] the traditional SQL queries. You will [1:58] need to compile with the manual queries. [2:00] You need to do the time consuming for [2:02] non-technical users. Technical people [2:05] they know like how to write queries. But [2:07] those who don't know it's very [2:08] beneficial for them. What will the [2:11] benefit happen? You will be having NLP [2:14] running on that. It will be realtime [2:16] query optimization like when you ask the [2:19] question it will give you what kind of [2:21] indexing what kind of optimization you [2:23] guys needed. And the last part is [2:25] enhance your productivity and decision [2:27] making like if you want to see only [2:30] three rows of data instead of like the [2:32] full data set you can ask on give me [2:34] only three columns what you looking for. [2:37] This is the AI integration workflow. If [2:40] you see user input natural language [2:42] query, there is a chatbot which is [2:45] talking to open AI GPD interprets and [2:48] generate the SQL query. There is a SQL [2:50] query execution and the realtime insight [2:53] delivered to the user. If you see in the [2:56] flow diagram, there is a DB schema. So [2:59] the only database schema is shared with [3:02] the OpenAI. Nothing else. Because if you [3:05] don't if openi don't know what query you [3:08] need or what data you need how he going [3:10] to generate a SQL query for you and then [3:14] there is the streamllet it is generating [3:16] a UI interface so you don't need any UI [3:19] developer to create a crazy UI and do [3:22] everything it's just like [3:23] straightforward Python codes you can do [3:25] it and then LLM chain if there is any [3:28] kind of [3:30] uh date columns are fetching from the [3:32] system machine learning model will kick [3:34] in immediately and generate the [3:36] predictive future for next 12 months how [3:38] the data will look like. I will show you [3:40] guys all the examples in the next slide [3:42] and then after everything is done SQL [3:45] execution happen you can able to see a [3:47] SQL query and all the optimization. [3:52] Okay. Yeah. Here is the example how GPD [3:55] works on this this scenario. This is the [3:58] chatbot. You can see on the top on the [4:00] left side of the menu you can see all [4:02] the tables which is connected right now [4:05] and you can just simply ask the question [4:07] can you generate a query for me for [4:09] these ones. It generated the query. It's [4:11] give you the analytics like right [4:13] immediately these kind of indexing you [4:16] need to improve your system and you can [4:19] able to see the data right away [4:22] and on top of that if you ask the SQL so [4:25] it chatbot can you delete this can you [4:27] update that can you insert the records [4:30] it is not going to do that it is [4:31] completely 100% secure [4:37] SQL generation logic examples of [4:39] generated here so here in The bigger [4:41] picture you can see uh give me the top [4:44] 10 high value customers. It's run it [4:47] generate the joins for you and it give [4:50] you the suggestion like you need this [4:52] kind of induction which is missing. So [4:54] it's already predicted like what you are [4:57] curing the data what you looking for it [4:59] and how the how the data will come to [5:01] you and how you can optimize it in real [5:03] time as well. [5:08] uh so like query optimization like [5:12] resource allocations and everything uh [5:14] I'm going to ask you guys one question [5:16] how you guys comfortable with writing [5:19] the subquery [5:21] it's pretty hard or it's pretty simple [5:23] just yes or no that's it [5:26] anyone [5:29] no h sorry [5:34] I can't [5:36] simple uh you must be engineer then [5:41] yeah so subquery is like you are writing [5:44] query on top of another query so for [5:46] example if I want like can I have a [5:48] second highest employee salary then you [5:50] need to write the subquery to exclude [5:51] the first one and give me the second one [5:53] or you can put the limit like get me the [5:55] second one but here you can ask directly [5:58] hey can you give me a top 10 but I want [6:01] to exclude those people from my list [6:03] it's going to generate the query for you [6:04] and give the optimization right away. [6:08] So this is here a sub query sub case [6:11] here if you can see it's join the tables [6:14] on customers and purchase and it's [6:16] provided like what are the indexing is [6:18] looking previously we just saw two [6:19] indexing now it's giving me seven [6:21] because I'm looking for larger data set [6:23] now and it is not only restricted to [6:25] tables only if anyone have SQL server [6:28] experience here they know like we can [6:29] create storage procedure we can create [6:31] views so it will also suggest you do you [6:34] need a view for these kind of data so [6:36] they don't need to call the table [6:37] directly and it generate the view for [6:39] you and then you can use that view to go [6:41] into your SQL server and write and and [6:43] paste it there and execute it. It will [6:45] generate the view for you as well. [6:48] Uh [6:51] the real world use cases here like [6:53] banking and finance real time customer [6:55] segmentation is required retail [6:57] analytics like instant sale insight. So [7:00] if anyone is looking for like last 5 [7:02] years of data they can ask the question [7:04] get the data download the data and do [7:06] the analysis there and when I [7:08] implemented the LLM model which is CAT [7:10] boost model which is having a 90.95% [7:14] accuracy which is like 95% accuracy [7:17] which will give you a very perfect uh [7:20] forecast for the future. [7:24] So this is the cat boost model example [7:26] here. If if in this if in your SQL query [7:30] is there any kind of data return which [7:34] is having a real time or there is a time [7:36] stamp though the LLM will kick in [7:38] immediately and will predict the future. [7:40] So this data set which I have is from [7:42] 2012 to 2013 in my database but 2013 to [7:46] 2014 this is like next 10 month 10 [7:49] months of data which is predicted but if [7:51] you look at the graph they all are [7:52] identical that means the prediction is [7:54] 95% accurate here. So you can use that [7:57] data like what is going to happen in the [8:00] next uh 12 or 10 months in your company [8:03] or in your sales or in your insights or [8:05] anything. [8:09] Yes. [8:15] >> No, no, this is not chat GPT. This is [8:18] LLM. So as I mentioned like when you [8:21] fetch the data from the query and charge [8:24] GPT is not fetching anything. is [8:25] generating a SQL query for you. So once [8:27] you fetch the data and and it's deducted [8:30] the code will check if the column is [8:33] containing any kind of time stamp then [8:35] the LLM will kick in immediately. They [8:37] don't need any uh or anything and that [8:41] and that point we can able to see the [8:43] LLM like how the prediction is look like [8:46] like we did the prediction on customers. [8:48] So how the customer data will look like [8:49] in next 10 to 12 months. Okay. uh back [8:53] best practice for implementation and [8:55] action preparing your SQL server for AI [8:58] integration selecting app appropriate JP [9:02] models I'm using open AI because I love [9:04] the copilot but there is so many open AI [9:07] models are available like Microsoft [9:08] copilot or we can go with llama or [9:11] deepse whichever you feel comfortable [9:13] but it has to be uh integrated and cross [9:17] check what you are developing here uh [9:19] security and compliance consideration I [9:21] consider all those things because I'm [9:23] not sharing any personal record in my [9:25] model in my working model actually with [9:27] the open everything is 100% secured [9:30] there and continuous performance for [9:32] monitoring and iteration like how you [9:34] can improve it all the SQL query just [9:36] you guys saw I ran the same SQL query [9:39] against the SQL server and I got the [9:41] same results so I can say like the model [9:43] is giving me 100% accuracy on the result [9:46] set [9:48] what is next coming up so I'm working on [9:51] the next module which will be the AI [9:54] agent on the SQL server which is [9:56] following the same pattern but this will [9:58] be like more interactive model which [10:00] will be more like how uh anyone like [10:04] even the kid can ask like how can I [10:07] implement how like a like a human kind [10:10] of like anticipating as a result because [10:12] some people not interested in looking at [10:14] the CSP or the data tables there what if [10:17] it is like I'm going to ask hi how's the [10:20] VJ DJ punar is performing in last two [10:22] years is going to give me in the simple [10:23] English but it's going to run against [10:25] all the uh uh in the back end against [10:29] all the data set and everything. [10:32] So this is the complete flow diagram I'm [10:34] working on it which will be available by [10:36] end of this year and the model I just [10:39] shared for the chatbot which will I will [10:41] be publishing on public for free it will [10:45] be available by end of this month. [10:49] uh what I learned or what you guys will [10:51] learn from this one AI integration which [10:53] is significantly simplify how SQL [10:57] interaction will work. It's like SQL [11:00] server on your fingertips. Uh practical [11:03] methods for robust and implementation is [11:05] there a clear business advantage and [11:08] productivity. So if anyone is manager [11:11] here, no hard feeling. But if manager [11:13] want to see some report, they don't need [11:14] to bug the engineer. They can just go [11:16] and ask the question and the report is [11:18] there immediately and you can download [11:20] it. It's a simple download button. [11:22] That's it. Nothing need no new no new [11:25] need things to do. If senior director [11:26] wants something which is not required by [11:29] manager, he can just go and do