NYU M.S. in Computer Science, 2026 · Distributed systems, databases, and applied AI
I build systems that connect transaction processing, storage, and reliable AI. I am seeking software engineering roles where I can work on backend infrastructure and AI systems.
- Machine unlearning for shared LLM systems: investigating how to remove targeted knowledge while preserving the usefulness of shared models.
- Deterministic replication and transaction processing (Go): my experimental system reached 310K transactions per second, with zero ledger inconsistencies observed in the evaluated runs. These are experimental results for that system, separate from the C++ Caracal/Felis fork below.
- Caracal/Felis (C++): adapting an upstream deterministic concurrency-control system to CloudLab d710, repairing distributed builds, and extending multi-node YCSB experiments.
- LLM evaluation: building reproducible checks for grounded answers, tool-use behavior, fallback paths, and guardrails. PawPal+ includes an offline evaluation harness and a model card documenting limitations.
| Project | What I contributed | Stack |
|---|---|---|
| Caracal / Felis | CloudLab d710 adaptation, non-PIE distributed builds, and multi-node YCSB benchmarking in a research fork | C++, Linux, CloudLab |
| DuckDB storage engine | DuckDB/Badger integration with hash-partitioned key-value storage, versioned records, and batched writes | Go, DuckDB, SQL |
| PawPal+ | RAG-grounded assistant with Claude tool use, scheduling, guardrails, and reproducible offline evaluation | Python, Streamlit, Claude |
Languages: Go, C++, Python, Java, SQL, TypeScript/JavaScript
Systems: DuckDB, Badger, Linux, Docker, Kubernetes, AWS
AI: LLM orchestration, retrieval-augmented generation, tool calling, evaluation
Observability: Prometheus, Grafana, Loki


