Aspiring LLM Engineer focused on making LLM systems measurably better: retrieval quality, evaluation, and reliability.
I like to build the evaluation first, then iterate until the numbers move.
QA system over 100 public-sector RFP PDFs (~7,500 pages): text, table, and GPT-Vision image parsing, with a hybrid Dense + BM25 retriever.
- Audited my own team's evaluation after the project: found that the "Advanced" run reused the Naive index (so table/image parsing was never measured), along with grader false positives and an uncapped-MRR artifact
- Now running a clean ablation (hybrid → +tables → +images), scored per question type
- Wrote ~79% of the final codebase in a 5-person team
LangChain FAISS BM25 OpenAI Streamlit
Generative-AI platform that turns a restaurant's food photo into Instagram-ready ad images and copy, then uploads them automatically.
- PM and model lead in a 4-person team; top contributor (117 commits)
- Built the image-generation pipeline: style-reference generation and mask-based background replacement
PyTorch FastAPI Next.js MySQL
Graph-theoretic study of the real YouTube network (SNAP data) against Erdős–Rényi, Barabási–Albert, and Watts–Strogatz models.
NetworkX pandas NumPy
LLM / ML · LangChain, OpenAI API, Gemini API, FAISS, PyTorch, Hugging Face Backend / Web · Python, FastAPI, TypeScript, Next.js, React, MySQL Tools · Git, Docker, Jupyter, Streamlit