VisualVault — Semantic Image Search & Digital Asset Management Platform

02 — Approach

Solution

The workflow below is a visual summary of the documented implementation; it is not a claim of additional system components.

1. 01

   Designed a multi-service ML platform from scratch: FastAPI + Celery workers for async inference, Qdrant for vector search, and a Streamlit UI — deployed via Docker Compose with GPU-aware worker containers

   →
2. 02

   Optimized CLIP inference end-to-end: exported to ONNX, compiled a TensorRT engine, and built a model registry to swap providers at runtime without restarting the API — 3.8× latency reduction in practice

   →
3. 03

   Closed the retraining loop automatically: YOLO detections get flagged, pushed to Label Studio for human review, written to Postgres, and Celery Beat triggers fine-tuning once corrections hit a threshold — no manual intervention needed

03 — Engineering judgment

Key decisions

• 01

  Closed the retraining loop automatically: YOLO detections get flagged, pushed to Label Studio for human review, written to Postgres, and Celery Beat triggers fine-tuning once corrections hit a threshold — no manual intervention needed

• 02

  Added embedding drift detection to catch distribution shift on incoming images before it silently degrades search quality; the OOD test set scored 2.3× above the alert threshold, confirming the signal was real

• 03

  Extended the platform from images to video — FFmpeg extraction feeding into the same async Celery pipeline, with a WebSocket live feed for real-time previews

04 — Evidence

Results

This project record verifies the delivered implementation, but does not report a comparable performance metric.

Product & System Views