DeskZen: Autonomous Workspace Ergonomics & Declutter Agent

DeskZen uses OpenCV 5 and AWS Bedrock (Amazon Nova Pro) to score workspace visual entropy, detect liquid/cable hazards, generate interactive triage cards, and verify decluttering in an agentic loop.

Comment

Inspiration

A cluttered workspace is easy to recognize, but knowing where to start can be surprisingly difficult. Which items should move? How can you create room for your keyboard and mouse? Is that coffee mug too close to your electronics?

For many people, decluttering is not just about cleaning. It is about overcoming decision paralysis and figuring out how to create a workspace that supports focused work.

We built DeskZen to turn that uncertainty into clear, actionable steps. Our vision is to make workspace organization more accessible by combining computer vision, spatial reasoning, and AI-powered recommendations. Rather than simply identifying clutter, DeskZen helps users understand their workspace, plan improvements, and visualize a more organized environment.

What it does

DeskZen is an AI-assisted workspace organization and ergonomic analysis tool that transforms a photo of a desk into an actionable decluttering plan.

Its key capabilities include:

  • Workspace perception: Uses OpenCV and YOLOv8-based object detection to analyze workspace images and identify relevant objects.
  • Spatial analysis: Evaluates object positions, workspace obstructions, available working space, and potential hazards such as liquids placed close to electronics.
  • Ergonomic guidance: Assesses workspace layout using estimated measurements and ergonomic principles to identify issues with keyboard clearance, mouse movement, and screen alignment.
  • Visual workspace planning: Creates a 2.5D workspace preview that allows users to visualize proposed changes while preserving the appearance of their original items.
  • Actionable recommendations: Organizes suggested changes into practical actions, such as relocating a mug, repositioning a phone, clearing the mouse area, or moving cables out of the way.
  • Before-and-after assessment: Supports reassessment of workspace images to help evaluate changes in organization and spatial conditions.
  • Exportable workspace blueprint: Provides a visual reference for the proposed arrangement and its associated recommendations.

The goal is not simply to make a desk look cleaner. It is to help users understand what needs to change, why it matters, and how to get started.

How we built it

DeskZen combines traditional computer vision, object detection, geometric reasoning, image compositing, and cloud-based AI.

OpenCV and computer vision: We use image-processing techniques for spatial analysis, edge detection, visual measurements, and image inpainting. YOLOv8-based detection helps identify objects and distinguish essential work equipment from items that may need to be moved.

Spatial reasoning and digital workspace reconstruction: We developed a 2.5D scene representation that separates detected objects from the workspace background. This enables a user to preview alternative arrangements without relying on a generative image model to recreate the entire desk.

Geometric placement: A placement process evaluates candidate positions for relocated objects against workspace boundaries and other objects. The objective is to reduce overlap and preserve important working areas, including wrist clearance and the mouse movement corridor.

AWS Bedrock and Amazon Nova Pro: Bedrock integration provides an AI reasoning layer that can use structured workspace-analysis results to help produce recommendations and organize the decluttering workflow. A deterministic fallback is also included to support the application when cloud AI access is unavailable.

Interactive interface: We built a Streamlit interface that brings workspace analysis, recommendations, visual relocation, and workspace comparison into one workflow.

AWS deployment: DeskZen is deployed on AWS EC2, making the application accessible through a web interface rather than requiring every evaluator to configure the project locally.

Challenges we ran into

1. Recommending useful changes rather than merely detecting objects

Detecting a keyboard, mug, or phone does not explain whether its position is appropriate. We needed to consider each object's role and location relative to the primary work area. We addressed this by using workspace anchors and spatial relationships to inform recommendations.

2. Preventing relocated objects from overlapping

Simple placement rules can move multiple objects into the same area. We developed geometric placement checks that account for existing objects, workspace boundaries, and protected working zones.

3. Reconstructing the workspace without inventing objects

Generative image editing can introduce visual elements that were never present in the original scene. Our approach uses object cutouts, background inpainting, and image compositing to create a proposed arrangement using the user's existing items.

4. Accounting for different camera angles

Workspace photos vary in perspective, lighting, and orientation. Estimating physical distances from a single image is therefore challenging. Our approach uses recognizable hardware dimensions and spatial calibration techniques where possible, while recognizing that perspective and unusual equipment can affect measurement accuracy.

5. Connecting digital recommendations to physical action

A visually appealing preview is only useful if the user can act on it. We designed the workflow around actionable recommendations and reassessment, helping users move from an initial diagnosis toward a more organized physical workspace.

Accomplishments

  • Built an interactive workspace analysis and decluttering application rather than a concept-only mockup.
  • Combined OpenCV-based image processing and object detection with spatial reasoning to make workspace organization more actionable.
  • Developed a 2.5D visual planning approach that lets users preview object relocation without regenerating the entire image.
  • Implemented geometric placement checks to help prevent relocated objects from overlapping.
  • Integrated AWS Bedrock and Amazon Nova Pro into the recommendation workflow.
  • Deployed DeskZen on AWS EC2 for browser-based access.
  • Designed the workflow around a complete user journey: observe the workspace, identify issues, recommend changes, preview a target arrangement, and reassess the result.

DeskZen demonstrates how computer vision can progress beyond recognition and become a practical tool for understanding and improving physical environments.

What we learned

We learned that useful computer vision applications require more than accurate object detection. The system must understand spatial relationships, distinguish essential objects from potential clutter, and translate its findings into actions people can follow.

We also learned the value of combining deterministic computer vision with AI reasoning. Image processing and geometric rules provide a structured basis for spatial analysis and placement, while a language model can help turn those findings into understandable recommendations.

Finally, we learned that a digital preview and a verified physical result are not the same thing. Meaningful verification requires evidence from the updated workspace, not just confirmation that a proposed arrangement looks correct on screen.

What's next for DeskZen

We plan to extend DeskZen from a photo-based workspace assistant into a more responsive system for physical environments.

  • Real-time workspace monitoring: Explore continuous visual analysis that can identify changing workspace conditions and provide timely suggestions.
  • Augmented reality guidance: Investigate mobile or WebXR overlays that show where objects could be moved in the real workspace.
  • Improved spatial calibration: Expand support for different camera angles, desk surfaces, and non-standard equipment to improve the reliability of spatial measurements.
  • Stronger physical verification: Develop more robust before-and-after comparisons using actual workspace images, with clearer evidence of which conditions have improved.
  • Physical AI applications: Explore how the same principles of visual perception, spatial understanding, and action planning could contribute to intelligent systems that interact with and organize physical environments.

Our long-term vision is to make DeskZen a practical bridge between seeing a problem in the physical world and knowing what to do about it, one workspace at a time.

Built With

Share this project:

Updates

Submission history