AI Enterprise SolutionsGPT & Claude

Eco Sort Ai

Our users are waste management companies, recycling facilities (MRFs), municipalities, and large businesses such as supermarkets, logistics centers, and food and beverage plants. They rely on manual sorting, which is slow, costly, and inconsistent. Workers must make repetitive visual decisions at high speed, so accuracy drops with fatigue. Mixed streams end up contaminated and recyclable material goes to landfill. Soft, deformable items such as plastic film, bags, and textiles are especially hard to handle. EcoSort AI moves sorting from manual labor to accurate, data-driven sorting. It analyzes a photo or camera feed of waste in real time, detects each item, classifies the material, shows a confidence score, and decides which bin the item goes to. Low-confidence items go to a “reject” bin for human review, so people move from repetitive picking to supervision and quality control. ReflexGrip is the robotic layer for soft objects. It adjusts grip force from feedback, so the item is held without being crushed or dropped. Better sorting gives cleaner recyclable fractions and less material sent to landfill. This can lower the carbon footprint of waste handling, because more material is recycled and less is landfilled or incinerated. We treat this as a potential environmental benefit to be measured in pilots, not a proven result.

Open the demo Video GitHub

Preliminary score

64.0/ 100Promisingaverage of 2 judges

Value for the user17.5 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI20.5 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing12.5 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility8.5 / 15
Data requirements, running costs and a clear next step.
Originality5.0 / 10
A meaningful difference from common solutions.

What the judges wrote

GPT-6 AstraPreliminary66/100
Strengths
Your documented image-to-classification-to-review workflow gives AI a meaningful role in a concrete waste-sorting task. Your README clearly separates inference from simulated robot actions and describes useful safeguards, including uncertainty rejection, operator corrections, and fail-closed object checks. Your acknowledgment of dataset leakage, label conflicts, and camera limitations shows valuable attention to reliability.
To improve
Publish actual evaluation results with sample counts, clean camera-session holdouts, failure examples, and accepted-item accuracy versus rejection rate, alongside a manual-sorting baseline. Reconcile the submission’s YOLO/Next.js description with the README’s MobileNetV2/Python and COCO-SSD implementation, and provide a traceable input-to-routing example. Focus the next pilot on one facility workflow, with throughput and contamination targets, measured inference latency, and per-item running costs; validate soft-grip benefits separately on hardware.
Claude Opus 5.5Preliminary62/100
Strengths
Your classifier routes uncertain items to human review rather than forcing a wrong sort. That is backed by explicit rules: a confidence of at least 75%, a top-two margin of 15 points, three-frame burst agreement, and a fail-closed COCO-SSD guard that blocks non-waste objects like phones and people. Your README is unusually honest. It states that the robot, sensors and multi-object detection are simulated or manual, that thresholds are uncalibrated, and that duplicates across train/test splits may inflate results. You also back the quality work with code: a dataset audit tool, a clean-split preparation script, an evaluation script with a confusion matrix and per-class F1, and pytest/Node test suites. The operator-correction history with JSON/CSV export and the EPA WARM-based carbon estimate add practical value for an MRF workflow.
To improve
Your submission text and README disagree on the model: the submission says YOLO with an 'other/reject' class, while the README describes a five-class MobileNetV2 and says explicitly it is not YOLO. Align these so judges and users trust your claims. Put the actual test numbers (accuracy, macro F1, the main confusion pairs) in your written submission, and compare them with manual sorting accuracy or throughput at a real facility. The track asks for an adoption metric, so name one, such as contamination rate or picks per hour, and include rough running costs per camera or line. Finally, collect even a small held-out set of real conveyor-style or local photos, which would show whether the Kaggle-trained model transfers.

The team

Elcan Məmmədov
developer
Rauf Fəthi Qovaşin
developer
Namət Nağıyev
developer
Nərmin Əlizadə
developer
Alican MikayılovLead
developer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: We used an open-source computer-vision model (YOLO / TensorFlow-based) for waste detection and classification, adapted for our waste categories (plastic, paper/cardboard, glass, metal, organic, other/reject). The sorting decision logic and confidence thresholds are our own. We did not train a foundation model from scratch. All models and libraries are used under their public open-source licenses.
Data: We used publicly available waste image datasets from Kaggle, plus our own test images of everyday waste items. No personal or private data was used. The datasets are used only to train, demonstrate, and evaluate classification, and they remain under their original licenses.
Components: The web demo is built with React / Next.js and standard open-source libraries, and hosted on Vercel. AI inference uses the open-source YOLO / TensorFlow stack. ReflexGrip (adaptive soft gripping with force and slip feedback) and the conveyor/robot setup are presented as a designed concept and simulated flow in the demo. Everything beyond the standard libraries was written by our team.
Built during the event: Confirmed

Quality testing

We tested the full demo flow: app loading, image upload/capture, waste recognition, classification, confidence display, sorting decision, and the main UI interactions. We ran many different waste examples, such as plastic bottles, paper, cardboard, glass, metal cans, and organic waste, to check that the system responds correctly to different inputs. What broke: some edge cases and less common waste types were not always classified correctly. This happened especially when the image was unclear, blurry, or poorly lit, or when the object was partially covered, crumpled, dirty, or sat among other items. Visually similar materials (different plastics, or clear plastic and glass) were sometimes confused, and soft or transparent items like film and bags had lower confidence. What we did: uncertain results (below the confidence threshold) are routed to a “reject / needs review” bin instead of being sorted wrongly. We also improved how results and confidence are shown and fixed several UI issues found in testing. Next steps are more local data, controlled lighting, multi-object detection, and a separate benchmark for soft objects.

Worth 20 of the 100 points.

Running it

# How to Use the EcoSort AI Website **Before you start:** There is no setup. You don't need to install anything or create an account. Just open the link in your browser. ## 1. New Analysis (Image Analysis) 1. In the "New analysis" tab, choose a photo of the waste item. You can drag and drop it onto the upload area or pick it from the file dialog. 2. Click the "Analyze" button. 3. The "Analysis result" panel on the right shows the predicted category (for example, Plastic), the model's confidence and the alternative probabilities. ## 2. Robot & Camera 1. Open the "Robot & camera" tab. 2. If your browser asks for camera permission, allow it. 3. The panel shows the model's suggestion for the item and the proposed robot action route. ## 3. Azerbaijan Map - The container map opens on an OpenStreetMap base. - You can zoom in and out and filter points by type. - The map is still a demo. A real container database is not connected yet, and any pilot points are illustrative only. ## 4. Analysis History - Previous analyses are listed with the image, category, model confidence and time. - History is saved in your current browser only. - You can export the list as JSON or CSV. ## 5. Statistics - The distribution of saved analyses by category is shown as a chart. - The chart updates as you run new analyses. ## 6. ReflexGrip Simulation - You can rotate and zoom the 3D model with your mouse. - Below it is a conveyor and ReflexGrip simulation. Choose a material and parameters, click "Start simulation", and use "Stop" to end it. - This is a software simulation of the proposed physical system, not a working hardware prototype. ## 7. Video / Objects - This tab contains the project concept video and video-analysis features. - The concept video is a visualization of a possible future system, not a demonstration of working hardware. ## 8. Test Report - It shows the model's test results: overall accuracy, per-category precision, recall and F1, and the confusion matrix. - A dataset audit appears below the metrics. ## Notes - Results are model predictions and may not always be correct. - The camera feature needs a device with a camera and browser permission. - History and statistics are stored in the browser, so clearing browser data will remove them.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.