2025
Aven Support
AI customer support platform for the fintech company Aven - a 3D talking avatar named Dave answers questions over voice or text, backed by Pinecone semantic search and GPT-4, and validated by a 50+ case evaluation harness.
Overview
Aven Support is an AI customer service agent built for the fintech company Aven. Instead of a static chat window, users talk to Dave - a 3D avatar with synchronized lip movements - through voice or text, getting instant answers without waiting for a human agent.
The backend is built around a retrieval-augmented pipeline: FireCrawl scrapes and indexes Aven's knowledge base, Pinecone stores the embeddings, and incoming questions are matched semantically in real time before being passed to OpenAI's GPT-4 for generation. The result is answers grounded in actual company documentation, not hallucinated generalities.
Quality wasn't left to chance - a 50+ test case evaluation framework measures answer accuracy across representative support questions, so regressions in the retrieval or generation steps surface before they reach a user.
On the frontend, Three.js drives Dave's real-time facial animation - mouth movements are keyed to the audio stream - while the Web Speech API handles both speech recognition and text-to-speech, making the interaction feel more like a conversation than a support ticket.
Highlights
- 3D talking avatar with real-time lip sync driven by Three.js and the Web Speech API.
- Retrieval-augmented generation: FireCrawl ingests Aven's docs, Pinecone stores and retrieves relevant chunks in real time, GPT-4 generates the response.
- 50+ test case evaluation framework benchmarking answer accuracy.
- Voice and text input - users can speak or type; Dave speaks back.
- Flask API orchestrates scraping, embedding, vector search, and generation in a single pipeline.
- Material-UI interface designed for desktop and mobile.