ComputerVisionPlatform
YOLO + FastAPI + React Operator Runtime
An active internal CV platform built around a shared YOLO inference core, a formally assembled FastAPI runtime, PostgreSQL-backed catalog services, and a React operator frontend designed for real deployment workflow, not demo.
Formal Runtime Composition
The core design decision was to treat the platform as a composed runtime rather than a pile of inference endpoints. That meant drawing a hard line between operator surfaces, route groups, shared vision services, and catalog persistence.
Where the Surface Is Strongest
Frontend Runtime
A React operator surface organized around the workflows that matter in practice: inference, model and source management, stream operations, and catalog visibility.
Backend Assembly
The FastAPI side is assembled through an explicit `create_app()` pattern so new route groups can be added without collapsing runtime structure.
Inference Core
Every inference path delegates to the same shared core, so model state, preprocessing, and route behavior do not drift apart over time.
Catalog Backbone
Detection history and model metadata live behind a PostgreSQL-backed catalog with migration discipline, which makes the environment reproducible for dev and ops.
The Decisions That Made It Hold Together
Single Runtime Truth
One process owns all route groups and all shared inference responsibilities. That constraint keeps debugging sane and prevents parallel model instances from silently diverging.
Formal Assembly Over Script Growth
The backend was treated as an application that needed composition, not as a collection of scripts that happened to run. That is what made the platform extensible.
Operator Workflow First
The UI is not arranged around technical subsystems alone. It is arranged around what an operator needs to inspect, configure, and trust during deployment.
Where This Work Connects Outward
The platform now connects into the broader writing and lab surfaces, so the runtime lessons can be read as engineering strategy instead of staying trapped inside the project page.
Connection Lifecycles in Real-Time Vision Systems
Reflections on a CV platform stability issue where the real problem was not the model, but the lifecycle of MJPEG streaming connections. Making the stream path disconnect-aware and non-blocking made shutdown smoother, reduced resource contention, and turned a messy latency symptom into a much cleaner systems story.
WritingMemory Is Not Chat History
Reflections on assistant memory as a systems problem rather than a storage feature. Conversation history, useful memory, and durable context are not the same thing: a system can preserve messages and still fail to preserve meaning or carry forward what matters.
LabAgentic Assistant for a CV Operator Platform
An active systems investigation into how an LLM assistant should sit inside a FastAPI + React + YOLO platform: not as a chatbot bolted onto the UI, but as a tool-using orchestration layer over inference, catalog, and stream services with approvals, memory, and runtime guardrails.
LabDisconnect-Aware Streaming in a CV Runtime
A platform investigation on why shutdown latency, camera ownership, and seemingly unrelated inference slowdowns can all originate from long-lived MJPEG connections.