Agentic AI in bringing visual order to public spaces

A national public-sector authority that is conscious of the problem of “visual pollution.” A loosely defined term which includes all examples of visual chaos, visible mess or any other element that is not aligned with Saudi Arabia’s standards imparted by Saudi Vision 2030.

The client needed an agentic AI system to handle the inflow of data from over 33.9 million potential users and to support the municipal team in managing cases.

Industry

Public Administration

Headquarters

Saudi Arabia

Company size

Undisclosed

Wedo Solutions, an established IT services company with a strong track record in software development, prepared a mobile app for end users to submit information about visual pollution alongside a panel where the client’s employees could see the data, analyze it and implement all actions necessary.

When the project required specialised Agentic AI expertise, Vstorm was brought in as a contributing partner responsible for the intelligence layer of the solution. Wedo Solutions’ engineering discipline in software delivery and Vstorm’s specialisation in production-grade Agentic AI systems made for a complementary collaboration.

Vstorm’s impact, the TL;DR:

• Vstorm joined Wedo Solutions as the agentic AI partner on a Saudi public-sector project tackling “visual pollution” under Vision 2030, building the intelligence layer on top of Wedo’s existing app and UI.
• Core deliverable: an Arabic-language conversational AI agent letting client staff query incident data, statistics, and geographic KPIs via natural language instead of manual dashboards.
• Key technical fix: moved from a single LLM doing everything (interpretation, entity resolution, SQL generation) to a more reliable architecture — deterministic lookup tools for resolving Amana (administrative region) names, and a constrained reference list for incident classification, eliminating hallucinated categories.
• Data sources include a citizen-reporting mobile app (photos/location of pollution incidents), with the system designed to ingest additional sources as they’re added.
• In development: a root-cause analysis agent (drills into KPI shifts to find which regions/incident types drive changes, using deterministic stats with AI interpreting results) and a trend-forecasting agent (Python-based time-series forecasting, AI directs the analysis rather than performing the math itself).
• Tech stack (DeepAgents, LangGraph) was dictated by the client’s strategy consultancy and Wedo, not Vstorm’s usual stack — framed as evidence of Vstorm’s ability to work within partner-defined environments.
• Project is live; positioned as shifting the client from reactive incident response toward proactive operational planning across all 13 Amanas.

Making cities more appealing with Agentic AI

IGraffiti, garbage visible on the streets or any other form of visual chaos can have an immense impact on society. The problem was recognized by The World Economic Forum, there termed “anxious cities,” which has a heavy impact on the wellbeing of inhabitants.

The problem scales from one’s personal surroundings to city-level, with reducing visual noise being proven to reduce stress and improve wellbeing.

A national public-sector authority is conscious of the problem, introducing the concept of “visual pollution.” A loosely defined term which includes all examples of visual chaos, visible mess or any other element that is not aligned with Saudi Arabia’s standards imparted by Saudi Vision 2030.

The client needed an agentic AI system to handle the inflow of data from over 33.9 million potential users and to support the municipal team in managing cases.

Wedo Solutions, an established IT services company with a strong track record in software development, prepared a mobile app for end users to submit information about visual pollution alongside a panel where the client’s employees could see the data, analyze it and implement all actions necessary.

When the project required specialised Agentic AI expertise, Vstorm was brought in as a contributing partner responsible for the intelligence layer of the solution. Wedo Solutions’ engineering discipline in software delivery and Vstorm’s specialisation in production-grade Agentic AI systems made for a complementary collaboration.

The challenge

By the time we arrived on the picture, Wedo Solutions had already delivered a working proof of concept for an AI system, including a beautifully designed user interface with plentiful options and styles for data gathering and visualisation. They reached out to us at Vstorm to take ownership of the Agentic AI side of the solution, providing our specialized expertise to enhance the application while working with their developers.

The key difficulty was in designing an agentic AI system capable of processes many variable data points gathered from diverse data sources, from which a multi-agent architecture could be established as scope and needs grew, and align all of this with the broader context of Saudi Arabia Vision 2030.

A further difficulty in delivering our solution arose in the strategic overview produced by a global strategy consultancy engaged by the client for the program. The consultancy’s identified strategy was purely business oriented and there was a need to transform and translate this strategy into practical technical capabilities in order to fully address Saudi Arabia’s Vision 2030. Luckily, our background in agentic transformation consulting equipped us with the expertise we needed to translate this strategy into

The solution Vstorm was first engaged to build was a natural language interface, an AI conversational agent powered by an Arabic language LLM through which the client’s employees can query live incident data from across Saudi Arabia. The broader application handles statistics, incident tracking, and geographic KPIs. Our scope was the intelligence layer that makes that data accessible through natural language to all of the client’s representatives, keeping the reports and their current status actionable and available through a single, unified platform.

Arabic text-to-SQL

The core challenge was translating conversational Arabic-language queries into reliable SQL. The initial approach assigned the entire process directly to the language model: interpret the question, resolve all named entities, construct the query, and return a result.

That breadth was the problem. The more responsibility concentrated in a single LLM call, the more surface area there was for error and the slower the response.

Inconsistencies in Amanas

One recurring failure point was Amana resolution. Amanas are 13 administrative regions in Saudi Arabia. Every query touching geographic data required the system to identify the correct Amana ID in the database. Arabic place names often exist in multiple accepted forms and the model would resolve them inconsistently.

Because this lookup underpinned the majority of queries, instability here propagated across the entire system. Vstorm engineers extracted this function into a dedicated deterministic tool, a reliable lookup that returns the correct ID without relying on the model to infer it. The result was faster query execution and consistent behaviour on what had been a common failure case.

The same pattern applied to incident classification. The platform tracks specific, predefined types of visual pollution, each assigned a fixed ID. Rather than allowing the model to interpret incident types freely, which produced hallucinated categories that did not exist in the database, Vstorm provided the model with a constrained reference list. The system now selects from what exists, not from what it constructs.

Handling multi-source data

The underlying data comes from multiple sources. One of them is a public-facing citizen-reporting mobile application, a public-facing application available in Saudi Arabia through which citizens can photograph or record instances of visual pollution and submit them with location data, prompting the client to dispatch a response team.

A public-facing citizen-reporting mobile application is one of several data sources the platform ingests; the system is designed to accommodate new datasets and additional sources as the client’s data infrastructure grows.

Next steps

The project is already live and running. Yet there are areas where our Vstorm engineers are working on additional improvements as project scope continues to expand.

Root cause analysis

A further agent under development addresses root cause analysis. When a KPI shows a significant shift, like a spike in reported incidents over two months, for example, the system does not stop at surfacing the anomaly. It works downward through the data: identifying which Amanas contribute most to the change, which incident types shifted most sharply, and where the highest concentrations of impact sit.

The AI identifies what to test and interprets what the results mean and the statistical computation is handled deterministically. It is an agent using AI as a tool, not as a substitute for rigorous method.

Trend forecasting

The third agent in development handles forward-looking analysis. The client’s teams can ask what a given KPI is likely to look like next quarter, and the system will work from historical time-series data to produce a forecast. This forecasting uses Python-based statistical methods, including correlation testing and trend decomposition, executed by the agent as part of its reasoning process. The distinction worth noting is architectural: the language model here acts as a data scientist directing a set of analytical tools, not as the analytical engine itself.

The agent identifies seasonality patterns, detects directional trends, and applies appropriate data science methods to project future values. This moves the platform from incident response toward operational planning, giving the client a basis for anticipating where resources will be needed, not only where problems currently exist.

Technology

The project runs on a technology stack selected by the strategy consultancy and Wedo Solutions, rather than Vstorm’s standard tooling, namely DeepAgents and LangGraph, utilizing our capability at Vstorm to work within partner-defined stacks, a provided feature of our engagement model.