AI-Powered Study Abroad Consultancy Platform
A retrieval-augmented conversational system for a study-abroad consultancy, answering applicant questions from the firm's own documents instead of a generic model's guesswork.
the problem
Replace this with the situation before the work started: who was answering these questions by hand, how long it took, and what went wrong when the answer was wrong.
what I did
Built and evaluated a RAG conversational architecture on Flask, with Pinecone as the vector store, Cohere embeddings for retrieval and Groq for generation.
Add the decisions here — why this retriever, what chunking strategy, what was tried and discarded. The evaluation numbers below are the strongest thing on this page; the reasoning behind them is what makes them believable.
worth noting
- 92% Precision@3 and 0.89 NDCG in retrieval evaluation
- Pinecone vector store with Cohere embeddings
- Groq for low-latency generation
where it landed
Retrieval evaluation reached 92% Precision@3 and 0.89 NDCG on the test set.