Now

Engineering Summer Analyst

2026 · may — aug

• Built and deployed a production-ready full-stack agentic platform used internally at Goldman Sachs for Anaplan model-sync analysis, reducing reviewer effort by 67% through AI-generated risk and downstream-impact summaries.
• Implemented Java/Spring Boot services and a React/TypeScript frontend with a hybrid parser for 20+ report section types, combining deterministic extraction with LLM-generated summaries, risk signals, and impact analysis.
• Built a production RAG pipeline on OpenSearch with BM25 across 4 model-context chunk types, persisting source reports in DynamoDB and generated insights in AWS S3 for durable, repeatable retrieval.
• Delivered a 10x speedup in AI enrichment by parallelizing per-section retrieval and LLM calls with Java virtual threads and semaphore-gated concurrency, with centralized authentication, rate limiting, retries, and audit logging.
• Won the intern hackathon.

Startup School

2026 · jul

• Two days in San Francisco with a hand-selected group of builders flown in by YC. Talks, demos, and small group sessions with YC partners.

Undergraduate Researcher

2026 · jan — now

• Building a continual customization framework for text-to-video diffusion. The goal is a model that learns a new subject without losing the visual and temporal behavior it already had.
• Prototyping architecture-level extensions to DreamVideo and ModelScopeT2V-style pipelines. Subject conditioning, adapter-based customization, and temporal modules over denoising U-Net video backbones.
• Measuring what breaks rather than what improves. Catastrophic forgetting, identity drift, prompt controllability, and motion consistency across sequential customization tasks.

Software Engineer Intern

2025 · jul — sep

• Owned the agentic market-research workflow behind Scend's M&A platform. It turns a market thesis written in plain English into structured research on growth drivers, key players, and acquisition targets.
• Built the production data-acquisition layer in Python, Node, and Firecrawl. It covers five private-market entity types and normalized 189K+ companies and 14K+ investors.
• Hardened the pipeline across 125K+ enrichments with structured-output validation, filtering, deduplication, caching, and Postgres persistence.

NSF REU Researcher

2025 · may — aug

• Built MAVUL, a multi-agent vulnerability detection framework where one agent analyzes C code, a second critiques that analysis, and a third judges whether either was right.
• Evaluated on 879 real-world C/C++ vulnerable code pairs from PrimeVul against GPTLens, VulTrial, and GPT-4o.
• Published at IEEE TPS 2025.