Supporting Saudi Arabia's Vision 2030 with agentic AI in bringing visual order to public spaces.

The challenge

A working proof of concept that needed an intelligence layer

By the time Vstorm arrived, WeDo Solutions had already delivered a working proof of concept with a well-designed interface for gathering and visualising data. What the programme needed next was an agentic system able to process many variable data points from diverse sources, structured so that a multi-agent architecture could grow as scope did, and aligned with the wider goals of Vision 2030.

A second difficulty came from the strategic overview produced by a global strategy consultancy engaged for the programme. That strategy was purely business-oriented and had to be translated into practical technical capabilities before it could be built against. Translating it was work Vstorm had done before, in agentic transformation consulting.

Our work

Arabic text-to-SQL, and taking work away from the model

The first system Vstorm was engaged to build is a natural-language interface: a conversational agent powered by an Arabic-language LLM, through which the authority’s employees query live incident data from across Saudi Arabia. The broader application handles statistics, incident tracking and geographic KPIs. Vstorm’s scope was the layer that makes that data reachable in plain Arabic, on one platform.

The initial approach concentrated too much in one call

The core challenge was turning conversational Arabic into reliable SQL. The first design handed the whole process to the language model: interpret the question, resolve every named entity, construct the query, return a result.

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

Amana resolution became a deterministic tool

One failure point recurred: resolving Amanas, the 13 administrative regions of Saudi Arabia. Every query touching geographic data had to identify the correct Amana ID. Arabic place names often exist in several accepted forms, and the model resolved them inconsistently. Because this lookup sat underneath most queries, instability there propagated through the whole system.

Vstorm engineers pulled that function out into a dedicated deterministic tool: a reliable lookup returning the correct ID without asking the model to infer it. Queries ran faster and behaved consistently on what had been a common failure.

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

Built for more than one source

The underlying data arrives from multiple sources. One is the public-facing citizen-reporting app, through which people in Saudi Arabia photograph or record instances of visual pollution and submit them with location data, prompting the authority to dispatch a response team. That app is one of several inputs: the system is designed to take on new datasets as the client’s data infrastructure grows.

Next steps

From incident response toward operational planning

The project is live and running. Two further agents are in development as scope continues to expand.

Root cause analysis

When a KPI shifts sharply, a spike in reported incidents over two months for instance, the system does not stop at surfacing the anomaly. It works downward through the data: which Amanas contribute most to the change, which incident types moved hardest, where the heaviest concentrations sit. The AI decides what to test and interprets what the results mean; the statistical computation itself is handled deterministically. It is an agent using AI as a tool, not as a substitute for method.

Trend forecasting

The second agent under development handles forward-looking analysis. Teams can ask what a KPI is likely to look like next quarter, and the system works from historical time-series data to produce a forecast, using Python-based statistical methods including correlation testing and trend decomposition. The distinction is architectural: the language model acts as a data scientist directing analytical tools, not as the analytical engine itself.

The agent identifies seasonality, detects directional trends and applies appropriate methods to project future values. That moves the platform from incident response toward operational planning, giving the authority a basis for anticipating where resources will be needed rather than only where problems already are.

Technology

Delivered on a partner-defined stack

The project runs on DeepAgents and LangGraph, a stack selected by the strategy consultancy and WeDo Solutions rather than Vstorm’s standard tooling. Working inside a partner-defined stack is a deliberate feature of our engagement model, not an exception to it.

Summary

Thirteen regions, one chat

With Vstorm’s contribution, the authority can address visual pollution faster and more reliably, with the whole reporting and status pipeline across all 13 Amanas reachable from a single chat interface. That cuts processing overhead and speeds the deployment of city restoration work. The agents still in development mark a shift from reactive incident management toward forward planning, strengthening the aesthetic identity of Saudi cities and the quality of life in them.