Projects

Personal projects & case studies

Everything here is from my public GitHub repos: the architecture is drawn from the code, and every key decision links to the file it comes from.

Agentic · LangGraph

Web Research Agent

An 8-node LangGraph agent on Gemini that runs web and news search in parallel, has an LLM judge whether the snippets are enough, scrapes chosen pages with Crawl4AI when they aren’t, and compiles a Markdown report with confidence levels.

  • LangGraph
  • Gemini 2.0 Flash
  • Tavily
  • NewsAPI
  • Crawl4AI
  • FastAPI
  • React
RAG · Reranking

ARE: Aviation Regulation Explainer

A Streamlit app that pulls Title 14 of the CFR as XML from the eCFR API, chunks it by Part and Section, and answers questions with OpenAI embeddings, Cohere reranking and a context-only prompt, showing the source sections.

  • LangChain
  • ChromaDB
  • OpenAI embeddings
  • Cohere Rerank
  • GPT-3.5 Turbo
  • Streamlit
  • eCFR API
Agentic · Human-in-the-loop

Interview Agent

A LangGraph interview loop that pauses with interrupt() for each candidate answer and persists state with a SQLite checkpointer. Gemini picks the next question from a role-specific bank, scores the answer and writes feedback.

  • LangGraph
  • Gemini 2.0 Flash
  • SQLite checkpointer
  • MongoDB
  • FastAPI
  • React
  • MUI
Fine-tuning · QLoRA

Counsel Assist

GPT-3.5 turns grouped CounselChat answers into four-part support plans; those 860 examples fine-tune 4-bit Llama 3.1 8B with LoRA via Unsloth and TRL, and the merged model is published on Hugging Face.

  • Llama 3.1 8B
  • Unsloth
  • TRL SFTTrainer
  • QLoRA
  • GPT-3.5 (labels)
  • Weights & Biases
  • Hugging Face