DeepCamera: an open-source AI camera platform with autonomous hardware-aware skill deployment and local VLM analysis

What it solves

DeepCamera provides a platform to add AI capabilities to security cameras while keeping all data processing local for maximum privacy. It eliminates the complexity of manually configuring AI models for different hardware by using an autonomous deployment agent to handle installation and optimization.

How it works

The system uses a pluggable "skill" architecture where each AI capability (like object detection or scene analysis) is a self-contained module. A desktop application called SharpAI Aegis manages these skills, using a local LLM to automatically detect hardware (NVIDIA, AMD, Apple Silicon, Intel) and install the optimal model format (TensorRT, CoreML, OpenVINO, etc.). Skills communicate via a standardized JSONL protocol, allowing different models to be swapped interchangeably without breaking the pipeline.

Who it’s for

It is designed for users who want to implement advanced AI surveillance⁴such as person re-identification, fall detection, or VLM-powered scene analysis⁴on their own hardware without needing deep expertise in ML deployment or CLI tools.

Highlights

• Autonomous Installation: An LLM-driven agent reads skill manifests and automatically configures the environment and hardware acceleration.
• Hardware Agnostic: Native acceleration for NVIDIA (TensorRT), Apple Silicon (CoreML), Intel (OpenVINO), and AMD (ONNX).
• Privacy-First: Includes a depth-map anonymization skill that replaces live video with spatial activity maps to hide identities.
• HomeSec-Bench: A built-in 143-test evaluation suite to benchmark the security performance of local VLMs.
• Broad Skill Catalog: Supports real-time YOLO detection, SAM2 segmentation, and integration with Home Assistant.