Input, Processing and Output (IPO) of Massive Metadata from Image Recognit...
summary
Built a pipeline that captures frames with OpenCV, prefilters locally to send only positive matches, and routes metadata through cloud engines to databases and dashboards, handling about 80,000 requests/day across four clients. Stack includes OpenCV → AWS Rekognition / Google Vision → PostgreSQL / InfluxDB → Node.js / React dashboards, with Nginx/GX proxies and memoized API caching. Implemented QT_face and QT_vehicle content models (plates, age, emotion, timestamps, camera and store IDs), added DB indexing and cache servers to cut latency, and anonymized stored data.
AI-written summary · about
chapters (14)
- 0:00Distributed software architecture
- 4:52Conteo’s mission
- 5:55Gas station analytics
- 7:22Shopping mall analytics
- 8:53Medical clinic security
- 9:36Backend and database performance
- 11:06Open source tech stack
- 11:48Camera to dashboard workflow
- 13:32Local processing benefits
- 14:12Metadata structure
- 15:42Data privacy measures
- 16:25Data mining platform
- 17:56Scaling plans
- 19:01Acknowledgments
related talks
48m(Unpacking) AI's Impact on System Architecture
20mRecoding our workflows to FOSS4G
01h 18mArchitecture Kata
22mFrom 3000ms to 300ms: Lessons learned from optimizing Tanstack ...
04mBuilding Applications for a Studio in the Cloud at Netflix
11mBuilding Real-Time Feature Pipelines
01h 40mBuilding machine learning pipelines that scale
32mPromoting a data driven culture in a world of microservices