metadata

title: ScouterAI
emoji: 👓
colorFrom: green
colorTo: gray
sdk: gradio
sdk_version: 5.33.0
app_file: app.py
pinned: true
license: apache-2.0
tag: agent-demo-track
short_description: >-
  The agent capable of identifying the best model among the entire HuggingFace
  Hub to use for your needs.

ScouterAI - The Vision enhanced Agent

Welcome to ScouterAI, my Agents - MCP Hackathon submission. This app falls under the track 3 : Agentic Demo. The goal of the app is to demonstrate the capabilities of agentic llm's combined with more "traditional" deep learning computer vision. LLM's (and VLM's) are great models when it comes to interacting with the user and understanding its queries but are not (yet) capable of a precise perception of the images presented to them. Computer Vision models like object detection or image segmentation models are tailored models to accomplish these tasks but require some engineering to wrap them and be user ready. The idea of the agentic demo is to provide powerful LLM with access to expert vision models like object detection or image segmentation models. The agent can fulfill precise perception task on any object present in the image : detection, location, classification, masking, counting, etc...

Overview

In this preliminary app, the agent is a CodeAgent provided by the smolagents framework. Its interface consists of a chat interface with example and a gallery which is used to display the agent's work. The agent is provided with a set of tools :

Task model retriever : a RAG tool which, given a task (object-detection or image-segmentation) and a query (car e.g.), returns a list of models with their model id and the list of classes it is capable of detecting/segmenting. The list if based on a curated dataset of all the models available on the HuggingFace Hub, returns the mo
Computer vision models : Any object detection and image segmentation models available of HuggingFace
Image processing functions : Resizing, cropping, ...
Image annotation functions : Label, bounding box and mask annotators

To complete a user request

Use-cases

Stack

Agent framework : smolagents LLM : Anthropic Compute : Modal