ServiceBot IA

A final-year project I designed and built entirely on my own at EMSI. Support teams answer the same questions repeatedly, with no 24/7 availability and no visibility into actual customer satisfaction. I designed an AI agent that can answer, analyse and escalate.

• Automate answers to frequent questions using an indexed knowledge base (RAG).
• Detect customer dissatisfaction through sentiment analysis.
• Generate personalised responses to customer reviews.
• Escalate cleanly to a human with ticket creation when needed.
• Provide teams with a multilingual analytics dashboard.

• Design and development of the Python / FastAPI backend: chat, sentiment, reviews, surveys and reporting endpoints.
• Implementation of the RAG pipeline: document indexing in Pinecone and LangChain orchestration with Google Gemini.
• Development of the sentiment-analysis module and review-response generation.
• Development of the React interface: chat, dashboards and administration modules.

• 24/7 FAQ chatbot: answers generated from the knowledge base via RAG (LangChain + Gemini + Pinecone).
• Sentiment analysis: detection of emotions and dissatisfaction in conversations.
• Review responses: automatic generation of personalised replies to customer reviews.
• Automated satisfaction surveys: NPS, CSAT and CES.
• Human escalation: ticket creation and tracking when the agent cannot answer.
• Multilingual admin interface (including Arabic with RTL support): dashboard, analytics, PDF reports.

• FastAPI backend structured into modules: chatbot, rag, sentiment, reviews, surveys, tickets, orders, analytics, reports.
• RAG pipeline: documents chunked, embedded and indexed in Pinecone; context retrieval then generation with Gemini (Flash and Pro).
• LangChain orchestration chaining retrieval, prompts and model calls.
• React / Vite frontend with Redux: real-time chat, dashboards and administration.
• Dedicated test suite (pytest) and automatically generated Swagger / ReDoc documentation.

• Python
• FastAPI
• LangChain
• Google Gemini
• Pinecone
• RAG
• React
• Vite

• Answer quality: context retrieval (RAG) must be precise to avoid made-up answers — tuning of document chunking and prompts.
• Real multilingualism: translated interface including Arabic, with full RTL layout handling.
• Controlled escalation: defining the thresholds at which the agent hands over to a human rather than risking a wrong answer.

• Working prototype tested locally: chat, sentiment, reviews, surveys, tickets and dashboards all operational.
• Real application screenshots shown below.
• Known limitation: operation depends on paid external services (Gemini API, Pinecone) — not publicly deployed.
• Known limitation: demonstration project, not connected to real customer data.