CS master's student at UIUC, graduating December 2026. Currently a software engineering intern at Microsoft; previously an SDE intern at Amazon and AWS.
I like problems where a system has to be correct, not just working — solvers whose speedups you can measure, LLM pipelines that can't quietly make numbers up, evaluation sets that are allowed to return bad news. Most of what's here is that instinct applied to a different domain each time.
Looking for new grad roles starting 2027 — software engineering, and research engineering / applied science where the work stays close to the systems.
combinatorial-auctions — Engineering faster exact solvers for the auction Winner Determination Problem. Branching guidance gives up to 11× over Gurobi's default on hard mid-size instances — and is slower on the largest one tested, which is in the README too.
approximate-vcg-manipulation — A game-theory study of the same auctions from the incentive side: time-limited VCG solvers open a profitable bid manipulation (overbid 5×, profitable 72–75% of the time at 5–10% of exact solve time) that collapses under more solve time. Traces it to a branch-and-bound anchor effect. Write-up →
tv-writer-ai — A workspace for writing a TV pilot with a language model, where output is a branch rather than an answer. Every step generates competing alternatives; promoting one folds it into the project's canon, which then constrains everything generated afterwards, so the model can't contradict a choice you locked in three scenes ago. Next.js, Prisma, Gemini.
poker-coach — A no-limit hold'em coaching system built on one invariant: the model explains, it never calculates. Equity, pot odds, EV and MDF come from a deterministic layer and are handed to the model as given facts; a separate checker fails any response that cites a figure it wasn't given. A right answer reached by fabricating numbers is still a failure. 872 tests.
rag-document-qa — Retrieval-augmented QA over private PDFs: OCR fallback for scanned documents, sentence-boundary chunking, ChromaDB retrieval, and a ROUGE-scored evaluation set so the quality claims are checkable.
linking-wikipedia-pages — BFS, iterative-deepening DFS and simplified PageRank over the SNAP Wikipedia link graph. C++, course team project.
Python · C++ · TypeScript · PyTorch · Gurobi · Docker · AWS