--- date: 2026-04-23 meeting: Enterprise Brain Scrum source: Enterprise Brain Scrum - 2026_04_23 20_00 PDT - Notes by Gemini.txt duration: ~43 minutes --- # Enterprise Brain Scrum — 23 Apr 2026 **Date:** 2026-04-23 | **Time:** 20:00 PDT | **Duration:** ~43 min **Source:** Full transcript ## Participants **RG** — Rajashekar G Mannam **YR** — Yeshwanth Reddy Yerraguntla **VK** — Vara Kumar Jagarapu _(joined briefly at close)_ --- ## Act I — Post-Demo Debrief: What's Working, What Isn't (~00:00 – ~00:19) RG shared his screen showing the landing page. YR acknowledged design feedback had improved and response quality was better, though still unreliable. The first concern raised was one carried over from the previous day's demo. **Internet vs non-internet data not visually distinguished:** > *"Which are coming from internet, which are not — it's hard to distinguish. If I am a customer and if I just log in, I have no idea what is what. We have to somehow make the user understand what is what."* > — YR _(~00:00)_ RG clarified that the first answer in a response sequence is typically data-grounded, not internet-sourced, but confirmed this was not communicated visually in any way. YR narrowed the scope: he wanted to address only the internet-vs-non-internet problem first, nothing else in that pass. --- ## Act II — Internet Data Quality Issue: Relevance, Not Just Source (~00:19 – ~00:21) RG added a second concern beyond source labelling. Internet-sourced content was sometimes appearing without clear relevance to the question being asked. RG gave an example — the system had surfaced "500 million grant material cost sustainability focus" as an early-warning driver — but the relationship between that external data and the internal project state was not clearly articulated. > *"Whatever coming from the internet, that should be more relatable to the question — it should be like benchmarking — like how the world is moving and how is related to that — that kind of relationship should we highlight."* > — RG _(~00:20)_ YR acknowledged three distinct issues had now been surfaced: (1) source distinction, (2) internet content relevance/benchmarking, and (3) smart visuals for single values or percentages in internet-sourced content. --- ## Act III — Confidence Score, Hallucination, Sources, and Token Usage (~00:20 – ~00:22) RG listed additional requirements he wanted to see on the UI: - **Confidence score** — displayed alongside each answer - **Hallucination signal** — RG treated this as effectively the same as confidence: "hallucinating or not, same — those two are same" - **Sources** — links or references to where internet data came from - **Token consumption** — show the user how much of their usage budget is consumed, e.g. "you are using 5% — remaining 95%" YR noted the token/usage idea was something "they discovered now" (likely referring to a design conversation the day before) and asked RG to retrieve the Fathom recording from the previous day's client meeting, which had been recorded outside Google Meet. --- ## Act IV — Architecture Bug: UI Agent Not Called on Follow-Up Questions (~00:23 – ~00:25) RG demonstrated a live issue: when a follow-up question was asked, the Chanaka agent and Tatastel agent responded but **the UI agent was not invoked**. As a result, no visualisations were generated for the follow-up — only the base text response from the orchestrator layer. > *"The Chanaka agent and Tast agent are there, when I'm asking a follow-up question the UI agent didn't call. So that's why there is no visualization here — and this visualization is suggested by the status — those things we need to change in the architecture — I mean the system prompt itself needs to be updated accordingly."* > — RG _(~00:23)_ YR noted this clearly and said he needed to sit with RG for one to two hours to properly understand the architecture before it drifted further: > *"I can't drift too far from this project. I'll also need to understand — I need to understand what is not working."* > — YR _(~00:24)_ They agreed YR would come to the office on Monday for an extended architecture review session. YR asked RG to teach him how to run the full Enterprise Brain stack locally. --- ## Act V — YR Gets the Stack Running Locally (~00:26 – ~00:38) YR shared his screen and RG walked him through setup: 1. Pull the branch RG had shared (Tata Steel UI changes). 2. Review the `.env` file — 19 untracked files noted; skills and design folders present. YR noted design files were being used for UX analysis, not production. 3. Configure environment variables: Google Client ID not required (username/password auth, not Google OAuth). `POG` (PostgreSQL) session secret, DB host, JWT, and other DB variables set. 4. `uv sync` in the backend, followed by `npm install` in the frontend. 5. `npm run dev` to start the frontend. YR ran into a minor issue with the model config (`high` mode — only core agent active, not the full pipeline). RG clarified this was expected — in `high` mode only the core orchestrator runs. YR asked a pointed architectural question: > *"My bigger concern is — why is the prompt not in a DB?"* > — YR _(~00:37)_ RG confirmed they hadn't shifted prompts to the DB yet. YR said he would create a fresh branch and work on moving prompts to the database. --- ## Act VI — Conversation History and Final Architecture Notes (~00:38 – ~00:43) YR reviewed the backend logs and traced a query's execution through the pipeline. He observed the three suggested follow-up questions being generated in parallel after the primary answer, sourced from the Tatastel agent. YR noted the response appeared all at once (no token streaming) — which was expected given it was a visualisation-based response. YR confirmed he was "happy" with the pipeline flow as observed, but flagged the remaining problems: - Prompt not in DB - Visualisation architecture needs study - UI agent not being called on follow-up questions YR planned to review over the weekend and said he might call RG informally on Sunday. The call closed with VK briefly joining and saying goodbye. --- ## Todos - [ ] Fix system prompt / architecture so UI agent is invoked on follow-up questions, not just initial questions | author: Rajashekar G | deadline: unspecified - [ ] Move agent prompts from Python source files to database | author: Yeshwanth Reddy Yerraguntla | deadline: unspecified (fresh branch, post-weekend) - [ ] Add confidence score display to UI (per widget / per response section) | author: Rajashekar G | deadline: unspecified - [ ] Add source attribution for internet-sourced content (links, ranked by relevance) | author: Rajashekar G | deadline: unspecified - [ ] Add token consumption / usage indicator for users | author: Rajashekar G | deadline: unspecified - [ ] Ensure internet-sourced content is explicitly benchmarked / contextualised relative to the question (not surface-level retrieval) | author: Rajashekar G | deadline: unspecified - [ ] Retrieve Fathom recording from 2026-04-22 client meeting and share with YR | author: Rajashekar G | deadline: unspecified - [ ] Architecture walkthrough session (YR + RG in office) — teach YR how to run full stack and review architecture | author: Yeshwanth Reddy Yerraguntla + Rajashekar G | deadline: 2026-04-27 (Monday)