SA Video Library
Searchable Intelligence from 20 Years of Medical Animation
Transforming 3,300+ professional medical animations into a searchable, scene-by-scene knowledge base
with AI-powered analysis and full-text search capabilities.
📖 The Origin Story
Scientific Animations (SA) has produced 3,300+ medical animation videos over 20 years for industry leaders including Pfizer, Johnson & Johnson, Abbott, Medtronic, Baxter, Amgen, and NeoGenomics.
The Vision: Build a searchable library of ALL videos with scene-by-scene breakdowns including timestamps, narration, on-screen text, visual descriptions, visual actions, and concept tags.
Goal: Search "T cell" and find every relevant scene across the entire 20-year library instantly. No more manual hunting through thousands of videos.
🏗️ Architecture Decisions
Video Analysis Engine
Gemini 3 Pro Preview - Native video understanding that feeds actual video to the model, not just frames
Search Technology
SQLite + FTS5 - Full-text search with advanced query capabilities
Backend
Flask (Python) - Lightweight, flexible web framework
Frontend
Vanilla HTML/CSS/JS - No heavy frameworks, maximum performance
Hosting
Mac Mini (localhost:8891) + Static demo on Cloudflare Pages
🧪 Fidelity Testing Results
Test Video: Herpes virus mechanism of action (2:27 duration)
Date: February 4, 2026
| Analysis Level |
Scenes |
Total Words |
Cost/Video |
Cost/1K Videos |
| STANDARD |
9 |
730 |
$0.049 |
$49 |
| HIGH |
29 |
1,163 |
$0.075 |
$75 |
| ULTRA (thinking) |
19 |
1,087 |
$0.069 |
$69 |
| HIGH+ACTIONS |
29 |
1,274 |
$0.086 |
$86 |
Winner: HIGH+ACTIONS - Most scenes, most searchable words, includes visual actions field (nearly 2x ULTRA output)
Key Insights
- Interval Testing: 3-second vs 5-second intervals showed diminishing returns (3% more words, 24% more scenes). 5-second optimal.
- Token Economics: Gemini tokens based on DURATION, not file size. 4GB uncompressed = same cost as 100MB compressed of same length.
🚀 What Was Built
The Main Application (Tasks 268-271)
- Multi-Agent Development: 3 sub-agents built simultaneously - Sonnet (backend), Kimi K2 (frontend), Opus (integration)
- Brand Alignment: Restyled to match scientificanimations.com with purple #2F135D, cyan #32C8E9, Khula font, glass morphism
- UI Preference: Scene detail view as table (not cards) per Zack's preference
- Status: Running on localhost:8891
Processing Queue System
Video Discovery
Found 3,298 MP4s + 3,316 JSON metadata files on external OWC SSD drive from SproutVideo
Smart Filtering
Detected 1,465 potential duplicates (same duration ± 0.5s + similar names)
Search Capabilities
4 search boxes with AND logic + exclusion support (prefix with -)
Queue Management
Drag-and-drop to queue, bulk operations, real-time processing log
Processing Pipeline
1
Download
1080p from SproutVideo
~10 seconds
2
Upload
To Gemini API
~28 seconds
3
Analysis
AI scene breakdown
~90 seconds
4
Storage
Save to SQLite
~2 seconds
Performance: Total ~2 minutes per video, ~$0.15 cost per video
Optimization: Using 1080p (88MB) instead of source (483MB) — same Gemini cost, 5x faster processing
Development Timeline
Feb 4, 2026 - Morning
Fidelity Testing Complete
HIGH+ACTIONS configuration selected as optimal
Feb 4, 2026 - Afternoon
Core App Development (Tasks 268-271)
3 simultaneous sub-agents build backend, frontend, and integration
Feb 4, 2026 - Evening
Processing Queue System
Video discovery, filtering, and queue management interface
Feb 4, 2026 - Night
Static Demo Deploy
Cloudflare Pages deployment with 3 sample videos
Task 273
Fire-and-Forget Queue
Server-side background processing, persistent queue
Task 277
Remote Queue Integration
Cloudflare Workers + D1 for remote video submission
Task 326
Table View Toggle (In Progress)
Adding table view alongside card view in library browser
📊 Current Status
4
Videos Processed
77+
Scenes Analyzed
3,496+
Searchable Words
3,298
Videos Ready
🔗 Live Deployments
Main Application: localhost:8891 (Mac Mini)
Static Demo: https://sa-video-library.pages.dev
Static demo includes 3 compressed videos with blob URL workaround for seeking
Sample Processed Videos
- Herpes Virus Mechanism of Action
- Biopsy Device Operation
- Lung Emphysema Pathology
- Precision Medicine Workflow
🎯 The Healthspan100 Connection
Challenge: SA client contracts mean final videos belong to clients. Can't use clips directly for new content.
Content Derivation Pipeline
1
Scene ID
Library search identifies medically accurate reference scenes
2
Keyframes
Extract via ffmpeg
3
Strip
Remove text, logos, drug/device references
4
Transform
AI style transfer (Nano Banana Pro) — color/camera/style changes
5
Generate
New 3-8 second video from transformed keyframes
🏰 Competitive Moat
3,300 videos × ~25 scenes = 82,500+ medically accurate reference scenes
No other health creator has this massive library of professional medical animations as an AI generation seed bank. Others start from text prompts → generic results.
Video Library + Healthspan100 = Symbiotic: Library built for SA business, content engine is a powerful byproduct.
🗺️ Future Roadmap
Batch Processing
Process all 3,298 videos
~$500 total, ~110 hours
Enhanced Search
Substring/infix search + semantic search with vector embeddings
Client Organization
Tag videos by client for internal organization and filtering
Healthspan100 Pipeline
Full automation of IP-clean content derivation
Team Access
Team-facing version for SA employees with role-based access
Remote Processing
Integrate existing Cloudflare Workers queue for remote submissions
💰 Cost Analysis
Full Library Investment
Processing Cost: ~3,300 videos × $0.15 = ~$500
Processing Time: ~2 min/video = ~110 hours sequential
Infrastructure: Mac Mini (owned) + Gemini API + Cloudflare free tier
Total Investment to index 20 years of work: ~$500
ROI Perspective: $500 investment unlocks searchable access to 20 years of professional medical animation content worth millions in production value.
🎉 Project Summary
From concept to deployment in one day - AI-powered video analysis meets enterprise search capabilities
20
Years of Content
3,300+
Videos Ready
82,500+
Potential Scenes
$500
Total Processing Cost
Ready for scale: Infrastructure built, pipeline optimized, competitive moat established.