Enterprise AI in Production

Description

Most enterprise AI demos stop at a polished chat response. This demo goes one step further: an AI assistant receives a natural-language business question, gathers relevant enterprise CRM context, and turns the answer into a live dashboard-style widget inside the conversation. The widget is not a static mockup or a pre-built report; it is generated from retrieved business data, streamed progressively into the UI, and persisted as a structured artifact that can be inspected after the conversation ends.

The session will focus on the production mechanics that make this trustworthy: tool use over enterprise data, source grounding, streamed block events, widget persistence, observability, token and cost visibility, and failure handling for generated UI. The demo is built around the question, “How do we know this dashboard is working, and not just looking plausible?” Rather than fully abstracting the system away, the live demo will hint at the internal trail: what data was selected, how the widget was assembled, where traces appear, and what happens when generated output should not be trusted.

What the live demo will show

A user asks a sales or revenue question in natural language, such as asking for pipeline movement, stage-level risk, or a compact visual summary of an account or opportunity. The assistant gathers the relevant business context, decides what kind of visual artifact would help, and streams an AI-generated widget back into the conversation.

The proof is visible in the demo rather than asserted in slides. The audience will see:

• The retrieved source data behind the widget.
• The streamed block lifecycle used to assemble the artifact.
• The persisted structured message content after generation completes.
• The final visualization and whether it matches the retrieved data.
• The observability trace with workflow steps, model calls, tokens, and cost.
• A failure or edge case where the system should not silently render an untrustworthy artifact.

Takeaways from session

1. AI-generated dashboards become production-ready only when they are grounded, observable, persisted, and failure-aware. The hard part is not making the model draw a chart; it is proving where the chart came from and whether it should be trusted.

2. Streaming AI artifacts need a different architecture from normal chatbot responses. Treating generated UI as structured, inspectable blocks gives teams a practical path from conversational AI to operational business workflows.

Which audiences is this session going to be beneficial for?

This session is especially useful for engineers, engineering leaders, AI product builders, and platform teams building agentic AI features inside enterprise software.

It will also be relevant for:

• Teams moving from chatbots to workflow-driven AI products.
• Data and analytics teams exploring AI-generated reporting.
• AI infrastructure teams responsible for observability, cost, and reliability.
• Product teams who need AI outputs to be explainable and inspectable.
• Architects designing tool-using agents over CRM, ERP, or warehouse data.

About me

I am an SDE 2 at Unravel Tech, where I work on agentic AI systems, RAG models, and production AI features for enterprise workflows. My work spans full-stack product engineering, AI orchestration, and data-grounded user experiences.

Outside engineering, I am also a writer, musician, and filmmaker, which shapes how I think about building AI systems that are not only technically reliable, but also clear, usable, and expressive for real users.

LinkedIn: https://www.linkedin.com/in/aviral-tuteja/