raksiam/nano-gpt-scratch

An end-to-end multimodal AI suite (PyTorch + Streamlit) that demonstrates engineering from raw tensor logic to a production-style, serverless-capable application. Intended for ML engineers, researchers, and developers who want a working example of custom transformer training, stateful assistant integration, and cloud-accelerated image synthesis. Key features: - Bard Engine: custom character-level GPT trained on Shakespeare; custom tokenization, multi-head causal self-attention, residuals, and sampling-based inference. - Assistant Engine: stateful multi-turn chat wrapper around high-throughput Groq cloud APIs with a memory buffer. - Image Studio: routes prompts to serverless clusters running a 12B FLUX.1 model to render high-resolution PNGs. - Repo includes dataset, model, train, generate, and Streamlit app; supports CUDA/MPS/CPU, AdamW training, and cross-hardware checkpoint loading. - Run: create venv, pip install torch streamlit groq requests, python src/train.py, streamlit run src/app.py.

AI-related keywords found in this repo's description, topics, or README summary — grouped by category. Each badge links to the corresponding ranking detail page.