# Minutes of Meeting — Manufacturing Demo Calibration
**Date:** February 27, 2026  
**Source transcript:** [Feb 27, 2026.txt](Feb 27, 2026.txt)

**Attendees:** Vijay Kamineni, Naveen Puttagunta, Satyasri Prabhakar Mantripragada

## Act I — Frame the lineage idea (00:00 – 05:38)

The conversation started with a hard calibration on how specific the manufacturing story should be. Naveen first framed lineage as a multi-stage production trace that could help explain faults, but Vijay immediately pushed back on turning that into a generic machine-sequence demo.

> *"No two plants will have the same machine sequence."*
> — Vijay Kamineni _(~00:03:30)_

That pushback changed the direction of the conversation. The risk was not technical feasibility; it was that the demo would feel like it only worked for one plant, one layout, or one customer. Vijay was explicit that manufacturing buyers usually see their own process as unique, which means a demo that overfits to one sequence can accidentally weaken the product story.

> *"We want to avoid that level of specificity."*
> — Vijay Kamineni _(~00:05:38)_

## Act II — Move from lineage to defect intelligence (05:38 – 10:29)

Vijay then steered the discussion toward defect classification instead of machine-sequence modeling. That was a better fit for the available synthetic data and a better way to show Enterprise Brain as an intelligence layer: count defect groups, identify repetitive failures, and show which defects are costly to ignore.

> *"We can always start talking about the possibility."*
> — Vijay Kamineni _(~00:09:32)_

The key idea was to shift from “what machine did this” to “what business problem does this defect create.” Vijay suggested adding cost, date, operator, and machine fields so the demo could surface useful patterns like night-shift concentration, Saturday spikes, or a smudge that matters much more than a scratch.

## Act III — Add business dimensions to the synthetic set (10:29 – 16:52)

The discussion then got more concrete. The team talked about defect groups, repeated defects versus one-offs, and whether the system could explain which defects should be targeted first because they are expensive or frequent. That framing makes the demo a business story about quality economics rather than a narrow computer-vision showcase.

Vijay kept emphasizing that the dataset can be synthetic, but the narrative must still feel operational and defensible. He suggested clustering by operator and time of day so the story has a believable shape without pretending to reflect a real plant record.

## Act IV — Keep the workshop story broad enough to sell (16:52 – end)

The close of the meeting focused on making the workshop narrative land with the client group. The team aligned on using a small but representative synthetic dataset, keeping the story broad enough to show capability across plants and customers while avoiding the impression that the system only works for a single machine topology.

The final shape of the story is: defect intelligence, not machine-line overfitting. That lets the team talk about cost, severity, reworkability, and operational patterns without getting trapped in a one-plant artifact that would be hard to generalize.

## Todos

<todo>
  Add cost, date, operator, and reworkability fields to the defect dataset so the demo can show business impact, not just classification.<br/>
  <span class="owner">Naveen / demo team</span>
  <span class="deadline">Before the next workshop</span>
</todo>

<todo>
  Reframe the manufacturing presentation around defect groups, repetitive failures, and expensive edge cases instead of machine-sequence generalization.<br/>
  <span class="owner">Vijay / Naveen</span>
  <span class="deadline">Next client review</span>
</todo>

<todo>
  Prepare one synthetic dataset that can support questions about scratch vs smudge severity and time-of-day / operator patterns.<br/>
  <span class="owner">Demo team</span>
  <span class="deadline">Before the demo dry run</span>
</todo>