Case 03 · Operations

Video understanding.
Operational quality.

Problem
Creator videos needed consistent feedback at operational scale.
Decision
Separated AI interpretation from code checks and revalidated recipients before sending.
Verified outcome
Delivered 1,109 feedback messages to 927 creators over two weeks.
1,109Feedback messages delivered in two weeks1
My role
System design · AI-assisted development lead · direct operation
Evidence
Verified operational records

Good feedback needs
more than generation.

The challenge was to connect video understanding and quality review into a useful workflow. This diagram summarizes that relationship at a conceptual level.

AI

Understand the video

Interpret frames and speech together, then draft feedback against the criteria I defined.

CODE

Check before delivery

Validate the message, compare the recipient record and block duplicate or incomplete outputs.

HUMAN

Set the scope of delegation

Review pilot results before expanding automated delivery, then handle exceptions in the operating report.

Components documented in operation

Python
Execution, checks & records
ffmpeg · faster-whisper
Video & speech preparation
Claude CLI
Context interpretation & feedback drafts

A stable target,
even when rows move.

An operating issue became a validation rule. This is an example of the decision process; private settings are omitted.

01 / ISSUE

Row positions changed

The source sheet changed between runs. A saved row position could no longer reliably identify the same video.

02 / DECISION

Reconcile by identity

Find the video by its stable identifier, then recheck the recipient and processing status immediately before delivery.

03 / VALIDATION

Verify the delivered result

Keep generation, delivery and result recording distinct. Count actual delivery from sending records; an AI draft alone is not a completed task.

What changes when rows move?

Synthetic illustration of record matching. This is not production code or a replay of operating results.

Current list · intended target: Video A

  1. Row 1Video B
  2. Row 2Video A
  3. Row 3Video C
Stored first-row position
Video B
Check the stable identifier
Video A · identity confirmed
No target is selected when the match is missing or ambiguous. This example only demonstrates record selection; it sends nothing.

Evidence from delivery.

Operational results and pilot validation represent different scopes of evidence.

9271

Creators reached

Recipients in the two-week operating sample.

1,1091

Feedback messages delivered

Messages actually sent during the same period.

552

Pilot assessments

Assessments included in pilot validation; this is not a measure of overall operational accuracy.

Sources and scope

  1. Work outcome summary and verified delivery records. Recipient and message counts cover the same two-week operating sample.

  2. Pilot assessment records. This validation sample is separate from operational delivery and does not establish overall accuracy.

  3. Operating and decision records support the responsibility split and the stable-identifier recheck. Detailed settings and private implementation are omitted. The delivery counts are not a measured effect of this individual change.