daily-media-update/

// daily morning briefing — GPT-4o writes a John Oliver-style monologue from your reminders, emails, and AI news. read aloud by OpenAI TTS.

// stack: gpt-4o · openai tts · kling ai · python

// status: in progress

next-word-prediction-lstm/

// next word prediction from scratch — lstm trained on medium article titles. autoregressive generation, one token at a time.

// stack: pytorch · nltk · pandas · numpy

// status: complete

finetune-t5-small-summarization/

// fine-tuning google t5-small on samsum dialogue dataset — seq2seq summarization with huggingface trainer api.

// stack: huggingface transformers · pytorch · datasets · accelerate

// status: complete

fine-tuning-gpt2-peft-lora/

// fine-tuning gpt-2 with peft-lora — only 0.24% of parameters trained. english quotes dataset, 4 minutes, one gpu.

// stack: peft · huggingface transformers · pytorch · accelerate

// status: complete

variational-autoencoders-mnist/

// conditional vae on mnist — generate any handwritten digit on demand. reparameterisation trick + kl divergence from scratch.

// stack: pytorch · torchvision · matplotlib

// status: complete

ai-anime-chatbot-with-rag/

// anime chatbot that scrapes any url, builds a faiss vector store, augments with duckduckgo search, and answers via groq llama 3.3 — every answer cites [rag] or [search].

// stack: langchain · faiss · huggingface · groq · duckduckgo · streamlit

// status: complete

quiz-generator-with-llama/

// paste any text, pick a difficulty — llama 4 generates 5 mcqs as structured json, rendered as an interactive quiz with auto-scoring.

// stack: groq · llama 4 scout · streamlit · python

// status: complete

self-rag-pipeline/

// self-evaluating rag pipeline — llm grades its own retrieved docs, checks for hallucinations, and validates its answer. stateful langgraph workflow.

// stack: langgraph · langchain · groq · chroma · huggingface

// status: complete