AI Engineer · Neptune Submerge × UoB
Prototype
Click an object in the scene (or focus it and use arrow keys) to raycast it and ask “what is that?”
?
Simulated MarineBot trace with stylized per-stage timings — the real pipeline ran Whisper STT → MiniLM + ChromaDB retrieval → LLaMA-3 → pyttsx3 TTS, triggered by Unity raycasts, fully offline and under 2 s end-to-end.
Highlights
- VR Voice Assistant: Served on a 2-person technical sub-team within a 6-member interdisciplinary squad to build "MarineBot," an offline, voice-enabled AI assistant for a Unity VR marine education game.
- RAG Implementation: Designed and integrated an information-retrieval RAG pipeline utilizing MiniLM embeddings and ChromaDB vector search coupled with LLaMA-3 to generate concise, domain-specific responses.
- Speech Pipelines & Inference: Integrated Whisper Speech-to-Text (STT) and pyttsx3 Text-to-Speech (TTS) pipelines alongside Unity raycasting for object-based querying, successfully implementing a modular architecture with real-time inference latency under 2 seconds.
- Technical Scoping: Co-authored a Technical Recommendations Report addressing project scalability, Quest hardware readiness, and optimized audio streaming.
- Client Delivery: Translated intricate user requirements into interactive communication flows, completing the functional prototype within a 6-week sprint and scoring a 4/5 client rating in Communication.
Technologies
- Unity
- MiniLM
- ChromaDB
- LLaMA-3
- Whisper
- pyttsx3