[ PIXEL :: QUEUE ]

VISION INTELLIGENCE INFRASTRUCTURE

A high-performance, dark-themed control panel for human-in-the-loop AI annotation, powered by decoupled ML microservices and robust task queues.

Architecture  •  Setup  •  Features  •  MLOps

---

⚡ Feature Matrix

PixelQueue embraces a strictly minimal, tech-forward aesthetic. It removes UX bottlenecks with pure speed, ditching heavy animations for instantaneous, DOM-optimized rendering.

Capability	Module	Description
Role-Based RBAC	Auth	First-party register/login with secure cookie sessions, plus project-level admin, reviewer, and annotator roles.
Asynchronous ML	Celery Worker	Non-blocking AI auto-labeling via PyTorch & Ultralytics integrations (YOLO/SAM).
Human-in-the-Loop	Review Queue	Imperative approval circuits. QA pipelines ensure 100% ground-truth validity.
Format Compilers	Export Engine	Distills annotation geometry into normalized COCO JSON and YOLO txt structures in seconds.
Zero-Latency UI	Canvas Renderer	Hardware-accelerated React-Konva staging. Zero-bloat OLED-optimized dark interface.

🏗️ Core Architecture

A highly decoupled, event-driven topology. The monolithic backend worker has been strictly partitioned into localized domains (tasks/, converters/) allowing infinite horizontal scaling of inference nodes.

graph TD
    %% Styling
    classDef client fill:#020617,stroke:#3B82F6,stroke-width:1px,color:#fff
    classDef proxy fill:#0F172A,stroke:#06B6D4,stroke-width:1px,color:#fff
    classDef backend fill:#0F172A,stroke:#8B5CF6,stroke-width:1px,color:#fff
    classDef worker fill:#111827,stroke:#10B981,stroke-width:1px,color:#fff

    Client[💻 Web Client<br/>React / Vite]:::client
    Nginx[🌐 API Gateway<br/>NGINX]:::proxy

    API[⚡ Core REST API<br/>FastAPI]:::backend
    DB[(🐘 Primary DB<br/>PostgreSQL)]:::backend
    Broker[[🟥 Message Bus<br/>Redis]]:::backend
    Storage{{🪣 Blob Store<br/>MinIO}}:::backend

    WorkerX[🤖 Inference Node A<br/>auto-label.py]:::worker
    WorkerY[📦 Compiler Node B<br/>export.py]:::worker

    Client -->|HTTP / WSS| Nginx
    Nginx -->|Proxy| API

    API -->|Read/Write| DB
    API -->|Dispatch| Broker
    API -->|Presigned URLs| Storage
    
    Broker -->|Consume| WorkerX
    Broker -->|Consume| WorkerY

    WorkerX -->|Read Weights| Storage
    WorkerY -->|Write Zips| Storage
    WorkerX -.->|Sync State| DB
    WorkerY -.->|Sync State| DB

Loading

🚀 Infrastructure Setup

Bootstrapping the entire constellation of microservices requires only Docker.

Important

Ensure ports 8000, 5173, 9000, and 5432 are open on your host machine.

# 1. Clone & initialize environment
git clone https://github.com/DhruvGarg111/PixelQueue.git
cd PixelQueue
cp .env.example .env

# 2. Compile and detatch all containers
docker compose up -d --build

# 3. Inject initial DB schemas, required buckets, and root users
docker compose --profile tools run --rm bootstrap

📡 Telemetry & Access

Intranet Target	Port Bind	Responsibility
Control Panel UI	localhost:5173	The primary frontend interface.
API Swagger Docs	localhost:8000/docs	Live OpenAPI schema for integration testing.
MinIO Console	localhost:9001	S3-compatible bucket explorer.

👤 RBAC Personas

Running the bootstrap script automatically provisions three demo profiles for local development and test workflows. Production onboarding is now self-serve through the register flow in the app.

Rank	Identifier	System Key (Pass)	Capabilities
[ADMIN]	admin@example.com	admin123	Global R/W. Cross-project administration, user assignment, pipeline triggering.
[REVIEWER]	reviewer@example.com	reviewer123	Queue isolation. Finalize or reject annotation tasks.
[ANNOTATOR]	annotator@example.com	annotator123	Self-serve project creation, canvas interaction, and ML-assisted annotation.

🧠 ML Operations

Located within the /scripts directory, an endogenous MLOps suite acts as the bridge between annotated data and deployment realities.

# 1. Sync & format remote datasets to local tensor inputs
python scripts/prepare_dataset.py

# 2. Execute GPU-accelerated YOLO bounds training
python scripts/train_yolo.py

# 3. Assess confidence vs. IOU against holdout splits
python scripts/evaluate.py

# 4. Serialize weights into the production registry
python scripts/register_model.py