Tata Steel: Use Cases Discussion Date: Mar 26, 2026 Source: https://docs.google.com/document/d/12sFgEi6RRHj05vzZ_jIpbznjOk9IRgmeTEQJHJjq6Wg/edit?tab=t.od431kymioa Invited: Satyasri Prabhakar Mantripragada, Yeshwanth Reddy Yerraguntla, Naveen Puttagunta, Vara Kumar Jagarapu --- SUMMARY Alignment achieved via three-phase M1 to M3 delivery plan for enterprise brain architecture, prioritizing reactive copilot and dynamic workspace. Data Agent Framework Readiness: The data agent framework skeleton is mostly ready for M1 delivery, enabling creation of specific entity agents. A missing component is a PostgreSQL facility for storing system prompts, which are currently hardcoded. M1 to M3 Milestones Defined: M1 focuses on the Reactive Copilot for question-answering with a redesigned UI, M2 on a dynamic workspace and agent training capability, and M3 on a fully integrated experience center. The decision was made to prioritize the full experience center (P2) over multi-data source orchestration (P3) in M3. Addressing AI Context and Quality: Development efforts must ensure platform independence for DUX by supporting multiple LLMs like Claude and Gemini. Developers are experiencing quality issues with Dakshi due to a lack of project context, necessitating a systematic way to capture all project activities. --- DETAILS Meeting Objective and Scope (00:00:00): Naveen Puttagunta initiated the sync-up with Yeshwanth Reddy Yerraguntla to review the current status, ensure alignment on sprint planning and deliverables, and discuss the business-level delivery vision for the enterprise brain architecture, which includes orchestrating multiple data sources for reactive Q&A, dynamic UI, and alerts. Yeshwanth Reddy Yerraguntla confirmed that nothing was omitted from the planned architecture (00:01:34). Data Agent Framework Milestones (00:01:34): Naveen Puttagunta requested that the technical plan be translated into business-focused milestones, noting that the team is delivering the data agent framework. The initial delivery of the data agent framework is a skeleton structure, and currently, training the agent for specific business use cases (like Tatastel or cautilia) must be done manually at the developer level, though the capability for the agent to be trained itself is planned for the next sprint (00:02:39). Data Agent Skeleton Readiness (00:03:42): Yeshwanth Reddy Yerraguntla confirmed the data agent skeleton is mostly ready, allowing for the creation of agents for specific entities such as cautilia or invictus. However, Yeshwanth Reddy Yerraguntla noted that certain good practices are missing, specifically the facility to store system prompts and instructions flexibly in a PostgreSQL database instead of hardcoding them into the Python code. Missing Data Source Agent Components (00:05:11): The "orange box," representing the database component that stores metadata like system prompts and instructions, is missing from the initial delivery of the data source agent. Yeshwanth Reddy Yerraguntla clarified that the component for the overall orchestration of the workflow is being stitched together, but the data source agent can still function temporarily without the external database by using certain strings embedded in the Python code. Real Time Decision Intelligence Core Agent Delivery (00:08:20): Yeshwanth Reddy Yerraguntla confirmed that the core Real Time Decision Intelligence agent is being delivered in M1. This agent currently functions as a hub for user-aware and data source agents, managing metadata like user identity and the question's mode (e.g., training mode, ad hoc question mode, or data refresh mode) (00:09:56). Future Real Time Agent Responsibilities (00:09:56): In the future, the Real Time Decision Intelligence agent's responsibilities will increase qualitatively, including synthesizing different sources of information and determining the correct data structure and method to store and retrieve organizational snapshots daily while respecting privacy and user roles. However, communication with other related boxes (two red, one pink) is currently missing (00:11:11). Reactive UI/Copilot Capability (00:12:15): Naveen Puttagunta confirmed that with the M1 deliverable, they will still be able to build an invictus agent and operate in a chatbot-like "reactive UI" mode where users ask questions and receive answers, similar to the manufacturing demos. The team agreed to use the term "reactive copilot" for this mode for internal reference. Reactive Copilot Functionality and Dynamic Content (00:15:13): Yeshwanth Reddy Yerraguntla committed to updating the description of the reactive copilot to include visual elements, such as charts with interactions and drill-downs, not just textual information. Naveen Puttagunta stressed the importance of ensuring the description does not suggest the system can only query a single data source but rather orchestrates multiple agents to synthesize an answer. Experience Center Terminology and M1 Delivery (00:17:35): It was agreed that the "dynamic workspace" is the method of output communication, which can be driven by either a reactive copilot or an autonomous experience center. Yeshwanth Reddy Yerraguntla confirmed they will receive the "reactive copilot plumbing" in M1, allowing for question answering capabilities, although the architectural databases are not yet fully set up (00:18:50). Reactive Copilot UI and Design Team Integration (00:21:16): Naveen Puttagunta noted that the design team is expected to deliver the reactive copilot redesigned UI and dynamic widgets (referred to as patterns) in code format. The expectation for Satyasri Prabhakar Mantripragada and Vara Kumar Jagarapu is to ensure the team integrates this redesigned UI, which is being developed in code rather than Figma (00:24:10). Connecting Visualizations to Enterprise Brain (00:26:32): Yeshwanth Reddy Yerraguntla raised a concern about connecting domain-specific visualizations (like those demonstrated for a manufacturing plant) to the enterprise brain framework. Naveen Puttagunta clarified that these complex visualizations should be viewed as patterns or widgets (e.g., a pie chart with segments, a heat map, or a trend line) that have associated metadata describing their capability. The engine must then decide which rich pattern to use based on the context (00:29:42). M1 Milestone (P0) for Reactive Copilot (00:32:27): The immediate M1 milestone (designated P0) is to deliver a reactive copilot with some base patterns and the redesigned UI for the Tatastel data source by the middle of the next week (Wednesday). This parallel work requires one stream to integrate the redesigned UI code and the other to prepare the invictus data and train the data agent (00:33:34). M2 Milestone (P1) for Experience Center (00:34:43): The M2 priority (P1), set for two weeks from the next day, is to deliver a dynamic workspace or "rudimentary experience center" that demonstrates the ability to dynamically generate UIs with rich patterns. Also included in P1 is building the capability to train data source agents on the fly, deferring complex features like separate trainer roles (arbback) to M3 (00:39:29). Proactive Experience Center Without Real-Time Data Refresh in M2 (00:36:12): For the M2 delivery, Naveen Puttagunta is asking for a proactive experience center, even without a full data refresh capability, to demonstrate the AI layer's ability to drive the UI dynamically. This can be simulated by linking and unlinking pre-prepared "snapshots" of static data to show how the dynamic dashboard changes (00:37:21). M3 Milestone (P2/P3) and Prioritization (00:43:42): The M3 focus (P2/P3), scheduled for about a month out (end of April), is on multi-data source orchestration and developing a "fully integrated experience center" that blends reactive and proactive modes. Naveen Puttagunta prioritized this full experience center (P2) over multi-data source orchestration (P3) and data refresh (P4), with the goal of having the core system working as envisioned by the end of M3 (00:49:13). Visualizations Framework Requirement (00:40:25): Yeshwanth Reddy Yerraguntla must analyze the patterns developed by Nihar from the design team and establish a UI framework where these rich patterns and widgets can describe themselves. This will allow the UI agent to consume the information and dynamically decide on the visualizations, which are currently being built using 3.js and D3.js (00:42:33). Discussion on Development Efforts and Storytelling (00:55:14): Yeshwanth Reddy Yerraguntla mentioned that the creation of the invictus report was an effort that came out of refactoring the capabilities of "Daksh," with a focus on good storytelling. Naveen Puttagunta and Yeshwanth Reddy Yerraguntla agreed to proceed with the priorities defined for M1, M2, and M3, while maintaining flexibility regarding data sets (00:56:36). Platform Independence for Development Tools (00:57:37): The team must ensure that the DUX methodology and related efforts are not solely dependent on Co-pilot and are adaptable to other large language models such as Claude, Codex, and Gemini. The goal is to merge current efforts so that DUX serves as a consistent layer for design, development, front-end, and back-end tasks, regardless of the underlying LLM or tool being used. Yeshwanth Reddy Yerraguntla is working on treating files as markdowns to make them accessible to both Co-pilot and Claude, although they do not have a formal solution yet (00:58:50). Streamlining Designer Workflow and UI Generation (01:00:03): There is a recognized need to further streamline the workflow for designers, particularly concerning how they generate documentation and UIs using tools like VS Code. The process should enable designers to generate documentation first, or generate the UI and then extract documentation, which is necessary for cleanup and development. The aim is to ensure UIs are built in a clean manner, incorporating component structures upfront, and allowing people to easily move between different systems when one system's tokens are exhausted (01:01:43). Training Session Review and Next Steps (01:02:45): The recent training sessions covered the necessary theory, and participants were able to identify their project goals, giving themselves one week to build them. The sessions focused on giving confidence regarding complex topics like an MCP server, ensuring attendees knew what it was and why it was needed, and referring them to documentation for the writing aspect. Naveen Puttagunta intends to follow up with Rama regarding the propagation of the training materials, which have been uploaded as five or six videos, to the next team (01:04:41). Trainee Enthusiasm and Hands-On Approach (01:06:46): Feedback regarding trainee enthusiasm suggests that while some individuals, like Rajes, are connecting the topics well, others may not be highly excited, potentially due to videos being turned off and general curiosity about implementation feasibility. Satyasri Prabhakar Mantripragada believes that hands-on experience and capstone projects will enable better relation to the concepts, while Yeshwanth Reddy Yerraguntla deliberately postponed the practical coding aspects (01:07:56). This approach was intended to allow trainees to appreciate the implementation challenges better after encountering roadblocks and failures during their initial application attempts (01:09:12). Addressing Dakshi's Narrative and Contextual Issues (01:10:45): Developers are reporting that Dakshi's usage is problematic, primarily because it either performs an excessive "info dump" or asks too many questions due to a lack of context. The core issue is seen as a process limitation where the system does not have sufficient project context to understand intentions, requiring a systematic way to capture all project activities, such as meeting discussions and salient points, on a daily basis (01:12:07). This comprehensive capture of context, including full meeting transcripts, will enable the AI to perform better, asking fewer questions and generating higher-quality output (01:15:00). Implementing and Retrospecting Agent Skills for Quality Output (01:16:26): To improve output quality, Yeshwanth Reddy Yerraguntla detailed the need for specific agent "skills," such as a "Storytelling mode" to ensure documents start with context for readers who have none, and a "Vasa skill" to ensure documentation is written correctly for Vasa rendering. A "Retrospective skill" was also created to automatically identify and fix mistakes in the skills used, such as the storytelling skill, so that errors are not repeated in subsequent conversations (01:17:57). Driving Operational Implementation and Methodology Change (01:19:18): A continuous architecture council, including Yeshwanth Reddy Yerraguntla, Gopalaru, Wenut, and Prabhakar, is needed to ensure that conversations and changes to DUX are actively translated into rollouts and adoption. Naveen Puttagunta emphasized that apathy, such as users not fixing a simple integration issue after a VS Code update, indicates that these changes are not being implemented operationally across all personnel. Furthermore, the team needs to drive changes in project estimation by transitioning from effort hours to complexity estimates, establishing a standard mechanism for evaluating and assigning story points for both UI and backend tasks at the scoping and development stages (01:20:52). --- SUGGESTED NEXT STEPS [Yeshwanth Reddy Yerraguntla] Update Copilot Doc: Update the communication modes document based on discussion notes. Ensure the Reactive Copilot section includes dynamic charts and user interactions. [The group] Grant Repo Access: Add Yeshwanth Reddy Yerraguntla to the UI pattern code repository. [The group] Share Nihar Prompts: Ask Nihar's team to share the UI generation prompts with Yeshwanth Reddy Yerraguntla for analysis. [Yeshwanth Reddy Yerraguntla] Build Agent Trainer: Deliver the capability for training data source agents on the fly as part of the M2 sprint goals. [Yeshwanth Reddy Yerraguntla] Connect Watana: Connect with Watana to find a better solution for LLM compatibility. [Naveen Puttagunta] Follow Up: Follow up with Rama regarding the plan to propagate training content to the next team batch. [Satyasri Prabhakar Mantripragada] Review Training: Revisit the missed core training session video. Try implementing the concepts. [Yeshwanth Reddy Yerraguntla] Drive Changes: Work with Gopal to drive necessary operational changes and methodology improvements.