NSTAD AI transforms official Djibouti statistics into searchable, traceable answers with intelligent table extraction and verified sources.

Inspiration

Finding reliable official statistics in Djibouti can be difficult. Statistical information is distributed across many PDF publications, annual reports, monthly bulletins, tables and different editions. A user may know the question they want to ask, but not which publication, page or table contains the answer.

We wanted to make this information accessible through a simple natural-language interface.

This is how we created INSTAD AI — Assistant Intelligent des Statistiques Nationales, an AI-powered assistant designed to help users search and understand official statistical publications from Djibouti.

What we built

INSTAD AI allows users to ask questions in natural language such as:

"Quel est le taux de chômage au 4ème trimestre 2025 ?"
"Donne-moi la valeur CAF des importations en février 2026."
"Quelles sont les exportations des grands groupes entre 2016 et 2020 ?"
"Compare les données entre plusieurs années."
"Quelle est la proportion des femmes dans la population ?"

Instead of simply searching for keywords, the system analyzes the user's intent, identifies the relevant statistical subject, period, metric and document, retrieves the most relevant evidence, and produces an answer with its official source and page references.

The main challenge

The biggest challenge was not simply connecting an LLM to a database.

The real challenge was making AI understand statistical documents correctly.

Many of the publications contain complex tables. During PDF extraction, information such as monthly periods, table columns, metrics and values can lose their original structure.

For example, a table containing:

2025 | 2026 | Volume | Valeur | Variation

cannot safely be treated as ordinary text. The AI must understand which value belongs to which column and which metric.

We therefore built a deterministic table-resolution layer that reconstructs and verifies table evidence before allowing the AI to generate an answer.

Our approach

INSTAD AI combines several layers:

Natural-language intent understanding
Document and publication retrieval
Semantic/vector retrieval
Statistical subject and period detection
Table reconstruction from extracted PDF content
Deterministic table and cell resolution
Relevance and hard-negative filtering
Grounded answer generation
Source and page attribution
Fallback mechanisms when reliable evidence cannot be found

The system is designed around an important principle:

If the system cannot establish reliable evidence, it should say that the information could not be determined rather than inventing a statistic.

What we learned

Building INSTAD AI taught us that reliable AI for statistical data requires more than a powerful language model.

Retrieval, document structure, table reconstruction, period normalization, metric identification and verification are just as important as the LLM itself.

We also created a golden evaluation dataset and use it to measure the system continuously instead of fixing only individual problems when users report them.

Our goal is to progressively improve the reliability of AI access to Djibouti's official statistical information.

What's next for INSTAD AI

We want to continue improving:

Multi-period and multi-publication questions
Complex statistical tables
Automatic verification of numerical answers
Confidence and evidence verification
Cross-publication comparisons
Multilingual interaction
Continuous evaluation across the entire statistical corpus

INSTAD AI is ultimately designed to become an intelligent gateway to Djibouti's official statistical knowledge.

Built With

ai
api
augmented
data
express.js
extraction
generation
javascript
llm
native
natural
node.js
ocr
openrouter
pdf
pgvector
postgresql
processing
rag
react
rest
retrieval
supabase
table
vector

Submitted to

AI Builders Hackathon

Created by

My contribution to INSTAD AI

Working on INSTAD AI was one of the most challenging and rewarding parts of our project. My main contribution was on the frontend and AI integration, where I helped transform a large collection of official statistical documents into an accessible, conversational experience.
I worked on building the AI chat interface, connecting it with our backend and designing the user experience so that users can ask questions about Djibouti’s official statistics and receive answers based on the available data and documents.
One of the biggest challenges we faced was making AI understand real-world statistical documents, especially PDFs containing tables, indicators, figures, and structured information. A simple RAG pipeline was not always enough: sometimes information could not be retrieved correctly, and sometimes the AI could generate an inaccurate answer even when the information existed in the documents.
This pushed me to think beyond simply “connecting an AI model.” I learned that building a reliable AI application requires a combination of good document processing, retrieval, structured data extraction, prompt engineering, validation, and a carefully designed user experience.

What I’m most proud of is helping turn a complex collection of statistical information into something much more intuitive: instead of searching through hundreds of pages, users can simply ask a question and interact with the data conversationally.

Through this project, I learned a lot about the real challenges of building RAG-based applications, working with large document collections, handling imperfect AI responses, and designing AI products where accuracy matters as much as the interface.

Most importantly, INSTAD AI taught me that building AI is not just about choosing a powerful model—it is about building the entire system around the model to make its answers useful, reliable, and accessible.

Moursal Ibrahim Moussa
Abdourahman Ibrahim