About
I build across the systems stack: production backend services, distributed job processing, LLM inference and GPU performance experiments, data/ML pipelines, native C++ tooling, and agentic AI applications. I care most about measurable performance, correctness, and understanding how the system behaves underneath the API.
Featured Projects
One representative project from each area; the category explorer above links to the broader set.
ML / Inference Systems
Distributed transformer inference experiments spanning vLLM serving, NCCL collectives, tensor-parallel communication, and PyTorch FX / torch.compile operator analysis on 2× A100 GPUs.
295.98 tok/s peak measured avg · 172.56 GB/s all-reduce · 252.87 GB/s broadcast
GPU / Native Performance
GPT-2-style byte-level BPE tokenizer implemented in C++ with a reusable thread pool and thread-local cache, benchmarked against tiktoken and Hugging Face Tokenizers.
2.11M tok/s at 8 threads · ~5× thread scaling · 91.6% of tiktoken throughput on the benchmark workload
Backend / Distributed
Flash-sale reservation backend with expiring seat holds and booking flows, using transactional PostgreSQL constraints to keep same-seat races correct under concurrent requests.
Concurrent same-seat race → exactly one hold succeeds · losing requests return 409 · Docker + CI/CD + Azure deployment
Data / ML Engineering
End-to-end batch ELT warehouse for the Olist e-commerce dataset, landing raw data in Azure Blob Storage, loading Snowflake, transforming and testing with dbt, and surfacing KPIs in Tableau.
Azure Blob → Snowflake RAW/STG/MART → dbt models + tests + docs → Tableau dashboards
Full Stack
Django/PostgreSQL marketplace for campus communities with listings, search and filters, favorites, seller messaging, meetup scheduling, reviews, moderation flows, and containerized deployment.
Agentic AI / RAG
Build-log triage agent that recognizes GCC/Clang, pytest, and TypeScript failures, extracts the first actionable error, and returns a compact fix plus rationale through the Agent Chat Protocol.
7.2K interactions · #2 most-used agent in the Fetch.AI track at CalHacks 12.0