Project Overview
The Problem
Website visitors previously had to navigate multiple static pages to find basic information about Mfano Bora Africa — services, attachment opportunities, ICT courses, careers, awards, contact details and application procedures. This was slow and frustrating. Many visitors left without finding what they needed. Customer care received repetitive questions. And there was no way to understand what visitors were actually asking about most often. Existing solutions were also brittle: simple FAQ-based chatbots broke whenever phrasing varied, while generic rule-based bots produced wrong or incomplete answers.
The Solution
We designed and deployed a production-grade AI chatbot with a multi-layer retrieval architecture: 1. Hybrid Retrieval Pipeline - FAQ semantic search (Sentence Transformers + keyword overlap) - ML intent classifier (TF-IDF + Linear SVM, 83 intents) - Disambiguation logic — only returns FAQ if confident AND top-1 beats top-2 by a clear margin - FAQ fallback layer for safety - Final generic fallback if nothing matches 2. Trained ML Model - Custom training dataset (~4,000 labeled examples across 83 intents) - Automated best-of-configuration sweep during training - Saved artifacts: model.pkl, vectorizer.pkl, label_encoder.pkl - Test accuracy: 84.6% 3. Verified Knowledge Base - 83 intent response entries in knowledge_base.json - 1,778 verified FAQ entries in faq.json - Chained follow-up questions guide users through topics naturally 4. Session-Aware Follow-Ups - Pending follow-up state stored per session - Handles "yes", "no", "tell me more" naturally - Carry-forward logic keeps the conversation on topic 5. Production Deployment - FastAPI backend running on a Windows production machine - Uvicorn server with hot-reload - ngrok tunnel for public HTTPS access - SQLite logging of every message (intent, confidence, source, FAQ ID, next intent) The chatbot handles ~80% of common queries autonomously, reduces customer care load, and continuously improves through conversation logging.
Target Users & Impact
Target Users: - Prospective students seeking attachment and internship opportunities - Graduates and early-career professionals exploring careers and mentorship - Businesses looking for ICT training, consultancy or partnerships - Award nominees, indus