Skip to content
DrepoTech
Back to the blog

Applied AI

GigaBin: Why Municipal AI Needs More Than a Smart Route

What our Cluj-Napoca prototype taught us about agentic workflows, explainable recommendations, and designing AI around real municipal decisions.

By 9 min read

Municipal waste collection can look like a routing problem from the outside: locate the bins, place them on a map, and calculate the shortest path. In practice, the decision is more complicated. Fill levels change, traffic changes, vehicle capacity is limited, service priorities compete, and an operator still has to understand why one collection is more urgent than another.

GigaBin began as a prototype during an international Erasmus Blended Intensive Programme at Babeș-Bolyai University in Cluj-Napoca. The project gave our team a concrete city context in which to explore an increasingly important question: how can several AI capabilities work together around an operational decision without turning the system into a black box?

A city does not need AI that merely sounds intelligent. It needs decisions that are useful, explainable, and possible to operate.

The real problem is coordination

A bin reaching a high fill level is only one signal. The system also needs to consider whether a collection vehicle has enough capacity, how far it is from the location, which other urgent bins can be served on the same route, and whether current traffic changes the practical order. Optimizing only one variable can create a recommendation that looks efficient on paper and performs poorly on the street.

That is why we approached GigaBin as a decision-support environment rather than a single prediction feature. Monitoring, prediction, urgency, routing, and vehicle assignment belong to the same operational chain. They can be separated into specialist responsibilities, but their output has to come together in one view that an operator can evaluate.

Why an agentic approach was useful

The term agentic AI is often used too broadly. For GigaBin, its value was specific: different roles could examine different parts of the decision. One role could interpret fill-level data, another could assess urgency, another could consider routes and traffic, and another could help coordinate vehicle assignment. A coordinating layer could then combine those contributions into a recommendation.

This separation makes a complex workflow easier to reason about. It also creates clearer boundaries for testing. If a route is poor, the team can investigate the routing assumptions. If urgency is inconsistent, the scoring logic can be reviewed. The architecture is not valuable because it imitates a team of people; it is valuable when it makes responsibilities, inputs, and failure points more understandable.

Prediction is useful only when it changes a decision

Predicting when a bin may become full is more useful than reacting only after it has already become a problem. But a prediction by itself is not an operational outcome. The platform has to translate that signal into a priority and connect the priority to a feasible collection plan.

In the GigaBin concept, predicted fill levels and urgency scoring help the system look ahead. Traffic-aware routing and vehicle assignment then move the idea closer to an actionable plan. Map data provides geographic context, while the operational interface gives the human user a place to inspect what the system recommends.

Explainability belongs in the product experience

For a municipality, an unexplained recommendation creates risk. Operators need to know whether a bin was prioritized because it is already full, because it is predicted to overflow soon, because it is close to another critical location, or because a particular vehicle can serve it efficiently. The explanation is part of the product, not a technical note added after the model has produced an answer.

This affects interface design. A useful operational screen should present the recommendation, the evidence behind it, and the important constraints without overwhelming the user. The objective is not to expose every internal calculation. It is to provide enough context for a responsible human decision.

A prototype should prove the decision loop

GigaBin was a concept and prototype, not a claim of a deployed municipal system. That distinction matters. A prototype should demonstrate whether the core workflow makes sense: can data become a priority, can that priority influence a route, can a vehicle be assigned, and can the resulting recommendation be explained?

A real implementation would require field validation, reliable sensor data, integrations with municipal systems, permissions, security controls, failure handling, and careful measurement. The prototype’s job was to make the larger system visible enough that these next questions could be discussed seriously.

From a Cluj-Napoca prototype to a platform for municipalities

Cluj-Napoca was the right place to make the idea concrete. The operational dashboard could work with a defined city map, real municipal context, suburb-level signals, collection vehicles, and a visible decision loop rather than an abstract smart-city presentation. Interest during the programme showed that the problem is not unique to one city.

The next direction for GigaBin is therefore broader but not generic. A municipal platform must be configurable around the operator’s fleet, zones, schedules, data availability, policies, and approval rules. The core product remains consistent—prediction, prioritization, routing, explanation, and learning—while the operating model adapts to the place using it.

The work was developed and presented by Ali Salihi of South East European University during the Erasmus+ BIP 2026 at Babeș-Bolyai University. Guidance from the international academic team, including Professors Abel Gómez Llana and Emilio Insfrán of the Polytechnic University of Valencia, helped create an environment where ambitious technical ideas could be tested against real-world responsibility.

What municipalities should ask before adopting AI

The first question should not be which model to use. It should be which operational decision needs to improve. Municipal teams should identify the people responsible for that decision, the data they can trust, the constraints the system must respect, and the evidence required before a recommendation can be acted upon.

They should also decide what happens when data is missing or contradictory. Can the operator override the recommendation? Is the reason for an override recorded? Can the system show uncertainty? These questions determine whether AI supports responsibility or simply adds another layer of complexity.

The lasting lesson from GigaBin

The most important lesson was not that agentic AI can be added to a city dashboard. It was that a credible AI product must connect intelligence to operations. Prediction, mapping, routing, and explanation become valuable when they help a real person make a better decision in a real context.

That principle now shapes how we think about applied AI at DrepoTech. Start with the decision, understand the system around it, make every automated role accountable, and keep the human operator close enough to judge the result.

Want this built for your company?

Use the article as a starting point. Tell us what you need, and we will map the delivery plan, timeline, and next step.

Start a project