Quantum Compute LLM Chat
A full-stack ML application showcasing end-to-end machine learning engineering, from custom model training to production deployment.
Training Infrastructure
- Oregon State University HPC
- NVIDIA H100 GPUs
- Two-phase training approach
- Phase 1: Book pretraining (17 epochs)
- Phase 2: Context Q&A fine-tuning (10 epochs)
- 480-test parameter optimization battery
RAG System
- 28,071 Q&A pairs in knowledge base
- Voyage AI embeddings
- Neon PostgreSQL + pgvector
- 100% retrieval accuracy
- Semantic similarity search
Tech Stack
- Backend: FastAPI, PyTorch
- Frontend: Flask, Vanilla JS
- Database: Neon PostgreSQL
- Embeddings: Voyage AI
- LLM API: Groq
- Custom Model: Modal GPU
- Hosting: Railway
Training Data
- 15,000 Claude-generated Q&A pairs
- Stack Exchange quantum computing posts
- Chain-of-Thought reasoning dataset
- 620K words from cleaned textbooks
- Rigorous data quality verification
Key Achievements
- 76% of users learned something new, 92% found explanations clear, 62% preferred it over general-purpose chatbots.
- Scalable cloud deployment, zero downtime over 6 months, sub-second inference, infrastructure under $5/month.