Enterprise AI in Production: Mumbai Call for Proposals

How to connect enterprise AI to documents, business data and existing software without sacrificing correctness, control or reliability.

Enterprise AI becomes difficult when users expect more than a plausible answer. A document question requires grounded retrieval; a sales question requires correct business definitions and reliable SQL; an operational request may require executing an action through an existing API. Routing all three through a generic agent hides fundamentally different failure modes. A system can retrieve relevant but insufficient evidence, generate valid SQL that answers the wrong question, or reason correctly and still fail when an external system rejects its action. The challenge is not simply making models smarter. It is designing the boundaries between reasoning, data access and execution.

At Sereno, our enterprise AI work with JSW Group spans conversational analytics, knowledge retrieval, website chatbot consolidation, and sales and dispatch analytics. I will examine how we approach these workflows using RAG, natural-language-to-SQL and existing APIs, supported by FastAPI-based backends and Azure infrastructure. The focus is on architectural decisions: which tasks require constrained execution, how to validate different outputs, and how to make common infrastructure reusable without forcing every workflow into the same abstraction. I will share the trade-offs involved, the operational signals we collect, and what I would measure differently to determine whether an agent is genuinely completing work rather than simply producing convincing responses.

The architecture combines a conversational interface with FastAPI-based application services and workflow-specific execution paths. Document queries use Azure AI Search and Azure OpenAI in a RAG pipeline; analytical queries use natural-language-to-SQL against structured business data; system operations connect to existing APIs. Supporting infrastructure includes Azure App Service, Cosmos DB, Key Vault, Microsoft Entra ID, GitHub Actions and Bicep/Terraform templates. Anonymous session and message events support usage analysis through Power BI. The principle is to share infrastructure while keeping retrieval, analytical correctness and action execution subject to different validation and failure-handling rules.

Across the wider enterprise programme, the environment has included more than 5,000 documents and 126 API integrations. A separate AI-assisted engineering workflow has involved approximately 1,000–1,500 engineering drawings. These figures cover different workstreams, not one agent handling all of them. Usage analysis includes anonymous session and message telemetry captured for reporting through Cosmos DB and Power BI. For the specific workflows discussed, I will distinguish verified production duration and measured outcomes from broader programme scale, sharing only figures cleared for disclosure.

A central trade-off is standardisation versus correctness. Reusing the same orchestration layer simplifies deployment, but retrieval, SQL and API execution have different failure modes. A relevant document is not necessarily sufficient evidence; a successful query can use the wrong business definition; and an API can fail independently of the model. I would invest earlier in separate evaluation sets for retrieval grounding, analytical correctness and end-to-end task completion, rather than treating answer quality as a proxy for reliability.

Attendees will leave with three practical takeaways: how to route tasks between retrieval, structured analysis and system actions; how to define correctness and failure handling for each path; and how to measure task completion, human intervention and error costs rather than response quality alone.

This session is intended for AI and software engineers, solution architects, data engineers and technical leaders bringing AI into production, particularly in manufacturing, banking, insurance and payments. Familiarity with LLM applications, APIs and databases is useful; experience with RAG or text-to-SQL is beneficial but not essential. My preferred format is a 30-minute talk, including questions.

I am Harsh Chaudhari, founder of Sereno Volante Software Services Pvt. Ltd. I lead product direction and delivery across enterprise AI systems, including conversational analytics, RAG, natural-language-to-SQL, application integrations and AI-assisted document workflows. Our enterprise work includes deployments for JSW Group. I can be reached at harsh@serenovolante.com or +91 8484926925. Website: https://serenovolante.com. LinkedIn: https://www.linkedin.com/in/harshchaudharisigma7/.

I am the founder of Sereno, which provides enterprise AI development and deployment services. This session draws on our implementation experience and may create awareness of our work, but it is intended as a technical case study rather than a sales presentation. Technologies discussed include Microsoft Azure services and AI orchestration frameworks. Customer data, confidential architecture and unapproved performance figures will not be disclosed. I am available to speak in Mumbai on 31 October 2026, between 11 AM and 4 PM IST, and to participate in editorial preparation. I understand that the session will be recorded and may be livestreamed for Fifth Elephant annual members.