# Minutes of Meeting - Enterprise Brain Security and Training Flow
**Date:** February 16, 2026  
**Duration:** ~17 minutes  
**Source transcript:** [Feb 16, 2026.txt](Feb%2016,%202026.txt)

**Attendees:** Yeshwanth Reddy Yerraguntla, Satyasri Prabhakar Mantripragada, Rajashekar G

## Act I - Security concern surfaces immediately (00:00 - 00:03)

The meeting opened with a concrete security concern. If the chat interface can surface table metadata, then it also has to resist prompts that try to mutate or expose data in unsafe ways. Satyasri asked whether a user could force DDL or DML behavior through the interface, and Yeshwanth answered with a first-pass guardrail strategy focused on query-level blocking.

> *"enterprise brain lo wherever we are dealing with table level queries queries should not contain alter drop keywords in blacklist JS."*
> — Yeshwanth Reddy Yerraguntla _(~00:01:38)_

The first minute made the larger issue clear: the product cannot expose internal details just because they exist. It has to protect the enterprise layer even while it is trying to answer questions.

## Act II - New tables need a human source of truth (00:03 - 00:07)

The conversation then shifted to onboarding new or obscure tables. Satyasri proposed a simple text-file style description per table so the agent can infer intent, while Yeshwanth described a trainer persona that can update system instructions directly when a new table is introduced.

> *"the trainer can now come to chat interface and say hey I just added these new tables. So this table means so and so that table means so and so please keep in your memory."*
> — Yeshwanth Reddy Yerraguntla _(~00:05:13)_

Both ideas converged on the same gap: the model needs a human-entered source of truth when the data model changes. The system cannot reliably infer an unknown table unless someone teaches it the meaning at least once, either through chat-based training or through a maintained metadata file.

## Act III - Entry points, controls, and cost (00:07 - 00:13)

The next stretch walked through the architecture diagram for entry points, guardrails, business services, and lower-level data sources. Satyasri pushed on access control, token usage, and how to explain the control model to clients; Yeshwanth agreed those business-facing controls matter even if they are not the top technical priority.

> *"how do I have control uh okay access layer standpoint aspect security aspect security access and control whom to give what and cost because end of the day everybody will start using left right and center shoot then."*
> — Satyasri Prabhakar Mantripragada _(~00:12:52)_

The key point was not just security, but operating control. The team needs a model that tells users how much they can consume, who can do what, and where the hard boundaries sit so the product does not become both expensive and uncontrollable.

## Act IV - Velocity becomes the checkpoint (00:13 - 00:17)

The meeting closed with a request to monitor velocity in the next week and use that as the next checkpoint for progress. The implication was clear: the security and training flow are now understood at a high level, but the team still has to prove that the guardrails, the trainer path, and the cost model can all be expressed cleanly in the implementation plan.

## Todos

<todo>
  Draft a checklist for query guardrails so table-level prompts cannot trigger destructive SQL keywords or internal data exposure.<br/>
  <span class="owner">Yeshwanth / Platform team</span>
  <span class="deadline">Next review</span>
</todo>

<todo>
  Define the trainer persona flow for adding new tables and updating system instructions or table metadata.<br/>
  <span class="owner">Yeshwanth / Satyasri</span>
  <span class="deadline">Next review</span>
</todo>

<todo>
  Decide whether obscure-table descriptions live in a text file, system prompt, or both, and document the update path.<br/>
  <span class="owner">Rajashekar / Platform team</span>
  <span class="deadline">Next review</span>
</todo>

<todo>
  Capture access-control, token, and cost-control requirements in the Enterprise Brain design notes.<br/>
  <span class="owner">Satyasri / Yeshwanth</span>
  <span class="deadline">Next review</span>
</todo>