Teknasyon AI Internship · Full-stack RAG · Voice
HP Chatbot AI
A full-stack RAG and voice product with guarded input, FAISS semantic retrieval, multi-source fallback routing, eight personality layers and dedicated voice output. I designed and built it end to end during my AI Department internship at Teknasyon, moved it from architecture to live deployment in roughly three weeks and reached more than 2,000 organic visitors.
01 · The problem
Why it exists
The challenge was not simply making eight characters sound different. Each response needed grounded context, safe input handling, conversation continuity and a reliable decision path when the primary knowledge base could not answer. The product therefore treats retrieval, routing and failure handling as core parts of the experience.
02 · Ownership
What I built
I designed and built the full product, including the RAG architecture, backend, retrieval and fallback system, character layer, interface, voice experience, visual direction and deployment.
03 · System
How it works
FastAPI coordinates a retrieval-first flow: input guarding, Sentence Transformers query embedding, FAISS semantic retrieval and context evaluation. When the primary knowledge path is insufficient, the system routes to Wikipedia, asks for clarification when the request is ambiguous or returns a controlled response when no reliable source is available. Qwen then combines grounded context with the selected character personality and conversation history before ElevenLabs produces the dedicated voice output.
04 · Evidence
What is proven
- Eight character personalities
- Eight dedicated voices
- Built in roughly three weeks
- More than 2,000 organic visitors
- FAISS, Wikipedia, clarification and controlled fallback routes