Umucyo

Overview:

In any modern government, procurement is one of the most vital, complex, and data-heavy operations. In Rwanda, public procurement accounts for a significant portion of government expenditure, with thousands of contracts awarded each year through platforms like Umucyo.

Yet, beneath this digital infrastructure lies a pressing challenge: the vast majority of procurement records are unstructured—inconsistent naming conventions, redundant item entries, and vague commodity descriptions make it difficult to analyze trends, monitor performance, or make data-driven purchasing decisions.

To address this problem, Épique AI launched SmartProcureAI—a research and development initiative aimed at transforming Rwanda’s procurement data into actionable intelligence. By leveraging natural language processing (NLP), machine learning, and real-time analytics.

Problem Statement:

The current procurement ecosystem in Rwanda suffers from data fragmentation and semantic ambiguity. The same item—say, “desktop computer”—may be entered in dozens of different ways across tenders: “HP PC,” “Dell Tower,” “Computer for Office Use,” or “Workstation (Intel i5).” Without a unified classification system, it becomes nearly impossible to:

• Track pricing trends over time

• Compare supplier performance across similar contracts

• Identify redundant or duplicated requests

• Forecast demand for recurring commodities

Procurement officers often rely on manual tagging and subjective interpretation, which varies from one institution to another. This leads to redundancy, inconsistency, and limits the government's ability to negotiate better prices or make strategic sourcing decisions.

Moreover, efforts to centralize this data are hindered by the absence of real-time classification tools that can standardize procurement entries as they are made—rather than retroactively trying to clean the data after the fact.

Our Approach:

To solve this, Épique AI built SmartProcureAI—an NLP-based system designed to bring structure, intelligence, and insight into public procurement data. The approach combined state-of-the-art machine learning with pragmatic tools for government integration.

1. Commodity Classification Engine

At the heart of the system is a BERT-based NLP model, fine-tuned to classify procurement items into a structured taxonomy of goods and services. Using a labeled dataset of simulated and anonymized procurement entries, we trained the model to recognize semantic patterns, normalize naming variations, and group similar items together.

For instance, the engine understands that “HP Laptop EliteBook” and “Portable PC for Project Managers” likely refer to the same commodity category. This significantly reduces redundancy and helps standardize procurement language across institutions.

2. Human-in-the-Loop Validation

To ensure reliability and improve learning, the system was designed with a human-in-the-loop workflow. Procurement officers can review, correct, and validate suggestions made by the model—feeding this feedback into an active learning loop that refines classification over time. This approach increases trust in the system while ensuring continuous improvement.

3. Dashboard and Insights Layer

We developed a visualization dashboard to help procurement managers view cleaned, classified, and standardized data in real-time. Key features include:

• Frequency analysis of commonly procured items

• Identification of redundant or duplicate entries

• Supplier distribution by commodity type

• Temporal trends in commodity pricing and demand

Key Outcomes:

The SmartProcureAI project delivered immediate and measurable improvements across several dimensions of procurement intelligence:

• 40% reduction in redundant entries across procurement records, based on before-and-after analysis in test deployments.

• Improved supplier selection through normalized commodity categorization, making it easier to compare offerings and pricing across vendors.

• Predictive analytics foundation: Clean, structured data opens the door to future capabilities such as demand forecasting, budget optimization, and anomaly detection (e.g., fraud or overspending signals).

Strategic Significance:

This case study illustrates the growing importance of AI for good governance. Procurement is not just about transactions—it’s about public trust, economic efficiency, and long-term national development. By structuring procurement data, Rwanda gains visibility into how taxpayer funds are spent and where savings or reforms are possible.

For Épique AI, this initiative strengthens its position as a leader in AI-driven public sector modernization—demonstrating that advanced technologies like NLP and machine learning can be responsibly deployed in complex government workflows.

More broadly, this work positions Rwanda as a continental pioneer in intelligent procurement systems. Rather than relying on imported solutions with rigid classification schemes, Rwanda is building localized AI tools tailored to its linguistic and institutional context—ensuring relevance, adaptability, and sustainability.

Conclusion:

Public procurement in Africa is often plagued by inefficiency, opacity, and manual workflows. But this is not a technical inevitability—it is a data challenge.

SmartProcureAI proves that with the right mix of machine learning, language understanding, and human validation, even the most unstructured data can be transformed into a strategic asset.

This project is not just about classifying items—it’s about unlocking better decisions, reducing waste, and elevating transparency in public spending.

As Rwanda continues to digitize government functions, structured procurement data will be the foundation on which future reforms, savings, and citizen trust are built.