# Enterprise Brain Scrum — 2026-03-26 **Meeting type:** Scrum / Architecture KT **Source transcript:** [20260326-scrum.txt](20260326-scrum.txt) **Google Doc:** [Enterprise Brain Scrum — Notes by Gemini](https://docs.google.com/document/d/1xDZepmYmrUJODBFISHf9OjypDpdNtMs1LNamfrKTco0/edit?tab=t.wwv4mcl5waau) **Participants:** Yeshwanth Reddy Yerraguntla, Rajasekar G, Pavhan Saii Gajjalakonda, Abhilash Adunuri, Amulya Maggidi, Kalakonda Harshith Rao --- ## Act I — Codebase walkthrough and agent structure (~00:00 – ~00:10) Yeshwanth opened with a walkthrough of the current codebase. The main code lives in the `chanaka` folder; running `make UI` is the single command that brings the full agent system up. The setup has four agents — two data source agents (Gmail and Jira), one real-time agent, and one user-aware agent — each running on a dedicated port. - Jira agent: port **8001** - Gmail agent: port **8002** - User-aware agent: port **8003** Each agent requires an `agent.py` and a Docker file, allowing individual agents to be deployed manually if needed. Both Gmail and Jira agents are currently running entirely on mock data. A "fake snapshot" is constructed by filtering this mock data based on demand and specific dates _(~00:07:17 – ~00:08:47)_. --- ## Act II — LLM strategy, mock data, and the A2A pipeline (~00:10 – ~00:22) The team is deliberately using a small local model — **qwen 3.5** — instead of Gemini Flash for the current debugging phase. The rationale is clear: > *"If the code pipeline works efficiently with a smaller model, it is highly likely to function well when we switch to Gemini Flash 3.0."* > — Yeshwanth _(~00:11:40)_ This forces every LLM tool call and prompt to be accurate enough to compensate for the smaller model's limitations. The cost-saving choice is a forcing function for engineering discipline _(~00:10:16)_. Mock data for emails, Jira tickets, and user preferences (including projects and dashboard settings) lives in a specific directory. For user context testing, the admin endpoint at `localhost:8010/admin` allows faking a user — for example, "Manisha" — which then reflects in user-aware agent responses, enabling queries like *"Who am I"* or *"What is my briefing today"* _(~00:12:51 – ~00:14:05)_. The **core A2A pipeline is operational**: the real-time agent can query the data source agents (Gmail and Jira), synthesize their responses, and handle generic queries. The pipeline already threads user identity through the call — the question is sent alongside metadata identifying who is asking. However, two controls are still pending: **training mode** and **ad hoc mode** toggles _(~00:18:20)_. One important gap was noted: the real-time intelligence agent currently **cannot update the org snapshot** because the whole system is still on mock data. All agents are running on the **pylog interface** _(~00:20:24 – ~00:21:51)_. --- ## Act III — Architecture internals and deployment details (~00:22 – ~00:40) The base agent model is defined in `base.py` as `enterprise_agent`, subclassing a **pydantic responder**. This gives two things: history retention across turns and compatibility with the pylog UI _(~00:26:32)_. The `A2A_context` component packages the following when the real-time agent calls a data agent: - Active user identity - Company / tenant details - Trainer flag - The question itself This propagates full context through the entire call chain _(~00:28:03)_. New agents follow the `new_agent_guide` documentation. Deployment is Docker-based per agent. One operational constraint: **downstream data source agents must be started before the main `chanaka` agent**. All agent URLs and ports are config-driven _(~00:32:01 – ~00:33:31)_. **Frontend decision made here.** The current dev version uses pylog, mounted at `/chat`. The system also exposes a WebSocket endpoint at `/ws` that can serve a proper frontend. The team agreed to move toward a single React app connected to the RT agent: > *"Future development should aim to use a single React application linked to the real-time agent, which will manage the data source agents in the backend."* > — Yeshwanth _(~00:35:40 – ~00:47:25)_ The working branch is `chanaka/alpha`. Model loading is conditional: uses an **MLX model** if environment variables are set, otherwise falls back to a light LLM _(~00:39:32)_. --- ## Act IV — Team assignments and close (~00:40 – ~00:47) The meeting closed with concrete assignments. The immediate goal for the day: > *"Connect the data to the system and ensure it is responding to questions in pylog by the end of the day."* > — Yeshwanth _(~00:46:12)_ Rajasekar G was asked to coordinate a follow-up call at **11:00 AM**. Pavhan was tasked with the data layer — reviewing the Vasa link and Drive files (Excel), understanding the schema, then dumping the full dataset into SQL. Abhilash and Amulya picked up React/UI integration. Harshith was directed to get a KT from Ganesh on the dashboards package _(~00:42:14 – ~00:46:12)_. --- ## Todos Commit and push latest code changes to the chanaka folder repository
Yeshwanth Reddy Yerraguntla 2026-03-26 Schedule the follow-up call — collect everyone's availability
Yeshwanth Reddy Yerraguntla 2026-03-26 Set up the next internal meeting call for 11:00 AM
Rajasekar G 2026-03-26 Review the shared Vasa link and Drive (Excel files) — understand data structures and DB schema
Pavhan Saii Gajjalakonda EOD 2026-03-26 Dump the full dataset into a SQL database based on the Vasa file structure
Pavhan Saii Gajjalakonda EOD 2026-03-26 Explore repo structure; determine how to integrate the React app via the /ws WebSocket endpoint
Abhilash Adunuri, Amulya Maggidi EOD 2026-03-26 Get KT from Ganesh on the dashboards package and determine integration path
Kalakonda Harshith Rao EOD 2026-03-26