Video Summarization Pipeline

JESUS_RENS — Video Summary & Rendering Pipeline

Authenticated video jobs are queued and processed through ASR, a local LLM, and ffmpeg rendering.

5 services30 routes225 testsLocal LFM2
Run Demo

Snapshot

Runtime
5 services
API
30 routes
Tests
225
Boundary
Gateway

Counts are based on compose.yaml and the current code structure.

Processing Flow

  1. GatewayBrowser requests pass through an authentication gateway.
  2. Job APIUpload and YouTube jobs are stored in the queue.
  3. WorkerASR, segment candidates, and summary plans are generated.
  4. Process 1Only transcription, summary planning, and segment generation are used.
  5. Renderffmpeg renders only approved outputs.
LLM output is constrained by a time_segments JSON contract, and generated final MP4 files are deleted from the server after 1 hour.
Public Demo

Analysis System Demo

Paste a YouTube URL to generate a fresh final summary video. When the job finishes, only the MP4 output is available for download.

  • Use a Protestant sermon video between 10 minutes and 2 hours.
  • The current training and validation data is limited to this category of sermon videos.
  • Public demo requests require admin-page approval for a 1-hour start or reservation window.
  • If another IP is already running a demo analysis, the estimated finish time is shown and you can reserve after approval.
  • Completed MP4 files can be downloaded for 1 hour after generation and are deleted afterward.

Sample source: 20260726 Suseo Sungji Church Sunday 1st Service

This Hour-
Today-
{}
Default summary selection
Progress 0% ETA -
Waiting for analysis

KPI & ROI

System

  • 5 Compose services
  • 30 HTTP routes
  • 225 test functions / 33 test files

Retention

  • Process 1-only analysis and rendering
  • Completed MP4 download window: 1 hour after generation
  • chat_only_mode and reuse_cache are disabled by default

ROI Assumption

  • Manual 180 min → 5-10 min after adoption
  • Estimated saving: 2.83-2.92h / KRW 169,470-174,455 per new video
  • Monthly scenario: 34.4-35.2h / KRW 2,057,567-2,105,418 for 6 new + 6 reprocessed videos

* ROI is an assumption until work logs are collected.

Basis: previous ROI formula and KOSA 2025 average hourly SW engineer wage of KRW 59,813. Manual and post-adoption times are scenario values before measured logs. KOSA source

Case Study

Authenticated Video Summary Job System

Problem

  • The analysis container needed a user access boundary that did not expose it directly.
  • Long videos make ASR timestamps, LLM output, and cut boundaries unstable.

Build

  • Separated Gateway, Auth/Admin, Job API, and Worker responsibilities.
  • Processed jobs through ASR → segment plan → rendering.

Result

  • Built a browser-operated flow with an operations boundary and job queue.
  • Separated final media from review cuts so weak validation is not overstated as a pass.
Evidence

Evidence

Architecture

Gateway/Auth/Admin/API/Worker boundaries and the /jobs contract are based on the design documents.

Validation

The page separates 3-case hard-gate pass evidence from the final-media strict-gate non-pass state.

Source

Based on compose.yaml, route/test code counts, and archi/docs design records.