Skip to content
YOLO-ToysArchitecture Whitepaper

One runtime for heterogeneous vision models, documented like a systems paper.

YOLO-Toys unifies YOLOv8, DETR, OWL-ViT, Grounding DINO, and BLIP behind a single FastAPI plus WebSocket service boundary. This site treats the repository as a technical artifact: architecture atlas, design essays, operational references, and research context — documented to the standard of a systems whitepaper.

Runtime architecture

YOLO-Toys is organized as a normalized serving runtime: transport-specific ingress, a central control plane, registry-backed dispatch, and model-family adapters that keep heterogeneous execution localized.

5 model familiesREST + WebSocketHandler / Registry patternLRU + TTL cache
<5msWarm latencyYOLOv8n on GPU, cache hit path
5Model familiesYOLO · DETR · OWL-ViT · G-DINO · BLIP
142Requests / secYOLOv8n cached, 20 concurrent users
85%Memory thresholdLRU eviction trigger for GPU safety
6Middleware layersSecurity → Metrics → Timeout → Rate → GZip → CORS
1Service boundaryAll model families behind one FastAPI runtime
Architecture blueprint
API surfaceHTTP REST + WebSocket ingress — routes stay thin, transport-specific, and replaceableTransport
Middleware stackSecurityHeaders → Metrics → Timeout → RateLimit → GZip → CORS in layered orderCross-cutting
Runtime coreModelManager — cache policy, concurrency guardrails, lifecycle ownershipControl plane
Handler registryHandlerRegistry — category inference, model metadata, deterministic dispatchDispatch
Execution adaptersYOLO · DETR · OWL-ViT · Grounding DINO · BLIP handlers — model-specific logic localizedExecution
Result normalizationStable public schema across all model families — same envelope for YOLO and DETRContract
Enter through architecture, operations, or research — each path is optimized for a different kind of expert reader.
The site is structured as a technical whitepaper with an academy and research wing. Every reading track is designed for advanced readers who want both the how and the why.

Released under the MIT License.