Orbit is an educational robot for kids built at NCT Robotik, where I work as an AI Developer. My job was the voice: making the robot listen, think and answer in real time — safely.
What I built
- Real-time voice assistant on top of the Gemini API and FastAPI.
- A fallback voice pipeline (STT → LLM → TTS) that I built and shipped to production, so the robot keeps talking even when the primary path is unavailable.
- Local STT, LLM and TTS services, so parts of the stack can run without depending on external APIs.
- A memory layer that lets the chatbot remember past conversations with a child.
- A profanity filter for the voice pipeline using Aho-Corasick and similar string-matching algorithms, catching inappropriate words fast enough for a live conversation.
- RAG systems and context caching integrations.
I also ran and monitored voice model training and LLM fine-tuning, and contributed to the web and mobile apps around the product.
Stack
Python · FastAPI · Gemini API · STT / TTS · RAG · Context caching · Docker · Linux