Retrieval-augmented generation pipeline over 1000 arXiv AI/ML papers. Combines sentence-boundary chunking, hybrid BM25 + semantic retrieval fused via reciprocal rank fusion, and a local cross-encoder reranker for relevance scoring rather than similarity proxies. A LangGraph agent decomposes compound queries, grades whether retrieved context is sufficient, and retries with reformulated searches when it falls short. Served through FastAPI on Hugging Face Spaces, with the reranker running locally.

Python · OpenAI API · ChromaDB · LangGraph · FastAPI · Flashrank · BM25

A lofi text adventure where an LLM acts as dungeon master, responding to natural language input with dynamically generated narrative, like talking out your own adventure novel. Features a simple combat system and a Pygame based CLI-style UI. The engine owns authoritative game state and applies changes only after Pydantic schema validation, preventing hallucinations. Context is managed across an injected engine state, a sliding hot context window with summarization compression, and a markdown file based cold storage for world logs and NPC histories retrieved on demand. An evaluation harness runs regression test cases against the LLM output contract after any system prompt change.

Python · OpenAI API · Pygame · Pydantic · Prompt Engineering

A working interpreter for a subset of OCaml, written in OCaml. Built a complete language pipeline with a regex based lexer, a recursive descent parser handling operator precedence and nested expressions, and a Hindley-Milner type inference engine using constraint generation and Robinson's unification algorithm. Built as part of CS314 Principles of Programming Languages.

OCaml · Recursive Descent Parsing · Type Inference

Full-stack web app that processes an uploaded resume PDF, analyzes a job description with Google Gemini, selects the most relevant experiences, and rewrites the resume for stronger alignment. Includes a validation layer that matches AI output against the original resume data to catch hallucinated content and fabricated metrics and forces output adherence to MIT resume standards. Built in 24 hours as part of HackRU Fall 2025 with two team members.

Next.js · TypeScript · Supabase · Google Gemini · React PDF