Building an AI-Powered Decision System for Smarter Waste Management
EchoLink is a WhatsApp-based AI waste management agent that connects street waste identification with responsible recycling. I developed a PyTorch-based vision classifier for waste categorization and a weighted decision model to guide decisions across the agent graph. My work focused on classification, hazard-aware routing, decision checkpoints, and reliable integration of AI predictions into actionable recycling workflows.
EchoLink: Building an AI-Powered Decision System for Smarter Waste Management
The Problem: Waste Management from the Customer’s Perspective
Street waste is not just an environmental problem; it is also a problem of accessibility, coordination, and decision-making. When people encounter discarded waste in public spaces, they often have no clear way to identify the right disposal method or connect with someone who can handle it responsibly.
From a customer's perspective, the problem can be understood through five key challenges:
1. Uncertainty About Waste Classification
People often struggle to determine whether discarded material is recyclable, non-recyclable, or potentially hazardous. Items such as batteries, electronic waste, broken glass, and mixed-material packaging require different handling procedures. Without reliable classification, users may dispose of waste incorrectly or leave it unattended.
2. Lack of Access to the Right Recycler
Even when someone wants to dispose of waste responsibly, finding a nearby recycler who accepts a particular material can be difficult. Different recyclers handle different waste categories, and customers may not know whom to contact or whether a recycler is available to collect the material.
3. Fragmented Reporting and Disposal Workflows
Reporting street waste, sharing its location, identifying its category, and arranging collection can involve several disconnected applications, phone calls, or manual steps. This friction discourages people from taking action, particularly when the waste appears insignificant or difficult to handle.
4. Safety Risks Associated with Hazardous Waste
Not all waste can be handled in the same way. Batteries, electronic components, and other potentially hazardous materials may require special precautions. Treating every waste item as ordinary recyclable material can expose people and waste handlers to unnecessary risks.
5. Lack of a Simple, Actionable Experience
Customers need a straightforward way to move from identifying waste to initiating its responsible disposal. A solution that requires technical knowledge, complicated registration, or manual coordination creates unnecessary barriers to participation.
The core problem, therefore, is not simply identifying waste. It is connecting waste identification with the appropriate next action through a workflow that is accessible, reliable, and safety-conscious.
The Solution: EchoLink and My Contribution
EchoLink is a WhatsApp-based AI agent designed to simplify the process of reporting street waste and connecting it with suitable recycling resources. Users can interact with the agent through text messages, upload an image of the waste, and share its location. Based on the available information, the system can classify the waste, evaluate its handling requirements, and initiate the process of identifying a suitable nearby recycler.
My primary contribution was developing the vision-based decision classification component and a weighted decision model that supports decision-making at the nodes of the agent graph.
Rather than treating image classification as an isolated prediction task, I approached it as part of a larger decision pipeline. The objective was to make the visual information useful to the agent's downstream workflow.
1. Developing the Vision-Based Waste Classifier
I built a decision-classifier vision model using PyTorch to classify waste images into the project's supported material categories:
- Paper
- Cardboard
- Plastic
- Metal
- Glass
- E-waste and batteries
These categories provide the visual classification layer needed to determine how a reported waste item should proceed through the system.
The model's output serves as an input to the broader agent workflow. Instead of relying entirely on a user's description, the system can use visual evidence to support its understanding of the reported waste.
An important distinction is that the classifier identifies a waste category, while the agent's decision logic determines what to do with that information. Keeping these responsibilities separate makes the overall architecture easier to reason about, test, and improve.
2. Building a Weighted Decision Model for the Agent Graph
My second responsibility was developing a weighted model that acts as a decision-maker at the nodes of the agent graph.
In a multi-step agent workflow, every node may need to decide whether to continue processing, request additional information, or direct the task toward a particular action. These decisions cannot always be made from a single model output.
The weighted decision model was designed to support decisions using the information available at the relevant checkpoints. Depending on the node, this can include the classifier's output, classification confidence, hazard assessment, user-provided information, and the status of required inputs.
Conceptually, the decision process can be represented as:
Input signals → Weighted evaluation → Decision checkpoint → Next agent action
This approach helps structure decision-making across the graph instead of allowing every stage to operate as an independent, disconnected component.
For example, an image classification result can help determine which waste-handling pathway should be considered. If the available information is insufficient, the workflow may need further confirmation rather than immediately proceeding. If an item is flagged as hazardous, the system must follow the appropriate safety-related path rather than treating it as ordinary recyclable material.
The weighted model provides a structured way to evaluate available signals, while the agent graph coordinates the subsequent actions.
3. Integrating Classification with Hazard Assessment
Waste classification alone is not sufficient for safe disposal. A material category does not automatically establish whether a particular item is hazardous.
For this reason, the classification output needs to work alongside a separate hazard assessment and the workflow's safety checks. For example, identifying an item as a battery-related object can trigger additional scrutiny, but the system must still apply the relevant hazard-handling rules.
This separation is important because a misclassification can affect more than the displayed label: it can influence the downstream disposal pathway.
The decision pipeline therefore needs to account for classification uncertainty, hazard-related signals, and the conditions required before the agent proceeds to the next stage.
4. Connecting Predictions to Real-World Actions
The broader EchoLink workflow uses the user's message, waste image, and shared location to support the identification of an appropriate recycler.
My work contributes to the intelligence layer that helps transform the image into a meaningful input for that workflow. The classification result informs the agent's understanding of the waste, while the decision model helps structure the next action at the relevant graph nodes.
This is the key architectural idea behind my contribution: a vision model should not merely produce a prediction; its output should participate in a controlled decision-making process that leads toward a useful real-world outcome.
Challenges: Production Checkpoints and Decision Reliability
Building a model that produces predictions in isolation is different from integrating it into a multi-stage agent system. In EchoLink, the important challenges extended beyond the classifier itself to the reliability of the decisions made throughout the workflow.
1. Maintaining Reliable Decisions Across Agent Nodes
Each node in an agent graph can depend on information produced by earlier stages. An incorrect or incomplete output can therefore affect downstream decisions.
A major engineering consideration was ensuring that the decision-making layer could evaluate the information available at each checkpoint and route the workflow appropriately. The system needed to account for missing inputs, uncertain predictions, and conditions that should prevent further processing.
This highlighted the importance of treating the agent graph as a coordinated decision system rather than simply chaining model calls together.
2. Handling Classification Uncertainty
Real-world waste images are not always clean, well-lit, or captured from a useful angle. Multiple materials may appear in the same image, objects may be partially obscured, and visually similar materials can be difficult to distinguish.
These conditions make classification confidence important. A model can produce a predicted class even when the evidence is weak, so a production workflow should not automatically interpret every prediction as a reliable result.
One important consideration was how classification outputs should be evaluated at decision checkpoints, including when the system should proceed and when additional user input or confirmation may be necessary.
3. Separating Material Classification from Hazard Decisions
A waste category and a hazard status represent different types of information. Conflating them can result in unsafe or inappropriate handling decisions.
The workflow therefore needs to distinguish the visual classification task from the hazard assessment and apply explicit safety checks before selecting a downstream path. Potentially hazardous items require particular care because the consequences of an incorrect decision may be more serious than an ordinary classification error.
4. Evaluating the End-to-End Workflow
A model can perform well on an isolated evaluation dataset while still producing problems when integrated into a larger application. In a multi-node agent system, the quality of the final outcome depends on how information moves between components and how each checkpoint responds to that information.
This makes end-to-end validation essential. The relevant checkpoints include whether the image produces a usable classification, whether the hazard assessment is applied correctly, whether the required user information is available, and whether the workflow selects an appropriate next action.
Testing these transitions is as important as evaluating the classifier's standalone performance.
5. Balancing Automation with Safety and Control
Automation is valuable only when the system knows when not to proceed. An uncertain classification, a potentially hazardous object, or incomplete user information may require a different path from a straightforward recyclable item.
The decision model and agent graph must therefore balance automated processing with explicit safeguards. Confidence-based checks, controlled routing, validation of required inputs, and confirmation where necessary help reduce the risk of propagating incorrect assumptions through the workflow.
These considerations reinforced an important lesson: production-ready AI is not just about producing accurate predictions; it is also about controlling how those predictions influence subsequent actions.
Making Summary
EchoLink demonstrates how computer vision and agent-based workflows can be combined to address a practical environmental problem. By allowing users to report street waste through WhatsApp, the system aims to reduce the friction between identifying discarded material and initiating responsible disposal.
My contribution focused on building a PyTorch-based vision decision classifier and a weighted decision model for the agent graph. The classifier provides waste-category predictions, while the weighted decision layer supports structured decisions at the workflow's checkpoints. Together with hazard assessment and the broader agent orchestration, these components help translate user-provided information into actionable disposal workflows.
The most valuable engineering insight from this project was understanding that reliable AI systems require more than a capable model: they need explicit decision logic, well-defined checkpoints, uncertainty handling, and safeguards around downstream actions.
Through EchoLink, I worked on the intersection of computer vision, decision systems, and agentic AI, applying these concepts to a real-world challenge where the quality of a model's output matters because of what the system does next.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article