A novel incremental RAG system called iRAG, which enables immediate interactive querying of large video repositories.

A hybrid approach that combines both lightweight and heavyweight VLMs where the lightweight VLM generates a text description, which then serves as input prompt for the heavyweight VLM under token limit.

LeanContext, efficiently extracts k key sentences from the context that are closely aligned with the query. The choice of k is neither static nor random; we introduce a reinforcement learning technique that dynamically determines k based on the query and context. Dynamic reduction of context size based on query contributes to the reduction of LLM API usage cost.

We proposed intelligent layer-selection for the secondary task while keeping the model for the primary task intact.

We developed FrameHopper,  an edge-cloud collaborative video analytics framework, that runs a lightweight trained RL agent on the camera and passes filtered frames to the cloud/edge server where the object detection model runs for a set of applications. 

We proposed MetaMorphosis in a split-multi-task-learning setup to add input obfuscation and task privacy to the overall system.