Projects
ArbiGraph
An open-source benchmark generator for testing whether tool-assisted LLM agents can retain, update, propagate, and discard typed computational state across arbitrarily scalable task graphs. Includes executable ground truth, controllable DAG topologies, exact evaluation, and released example datasets. ArbiGraph also provides reinforcement learning with verifiable rewards (RLVR) environments with dataset-backed and on-demand episodes, hidden verifier state, exact binary rewards, and per-node diagnostics.
[arXiv] | [GitHub] | [Hugging Face]
arXiv Research Agent
An evaluation-first retrieval-augmented generation (RAG) system for grounded literature review. Built with LangGraph, LangChain, Gemini, Chroma vector storage, BM25, hybrid retrieval, cross-encoder reranking, Pydantic, SQLite, and PyMuPDF. The project includes hand-labeled retrieval and screening datasets, reproducible evals, citation-grounding checks, resumable agent workflows, and 163 offline tests in CI.