Production systems with real users: three engineering internships, a nonprofit's entire platform as CTO, and fixes merged upstream into vLLM, Keras, Grafana, and the Kubernetes inference-routing SIG.
Software engineering is the core. Everything else is a stop along the way — each one a place you've just flown through.
Shipping production features across three engineering internships — Doro, HireBuddy, and BU Spark!. CTO and sole engineer of an international 501(c)(3)'s system of record, live across 20+ chapters in 9 countries behind a ~300-endpoint REST API. TerrierTracker serving 1,000+ students, as sole developer. The design work counts too: I authored the architecture and phased rollout plan for BU Spark!'s proposed unified database — an Airtable/Postgres hybrid sync feeding a portfolio generator, CRM, and project tracker — through three adversarial design-review rounds, recording which flagged risks were fixed and which were rejected as over-engineering.
68 pull requests across 25 repositories in 18 organizations — including vLLM, llm-d, KServe, Grafana, Datadog, Redis, Keras, VS Code, and the Kubernetes SIG for inference routing — 28 merged, 16 of them reviewed and approved by core maintainers including the project lead. A recent fix landed in Grafana core, shipping in 13.3.0: removing a tag from an annotation never unlinked it, because a column list forced a zero-valued ID into the DELETE predicate so it could never match. Each fix ships with the regression test that proves it. In llm-d I also merged the Batch Gateway's observability into the project's shared monitoring stack: 3 Grafana dashboards retargeted to the Prometheus datasource variable they lacked, plus 5 Prometheus alert rules kept namespace-scoped because the component's metric names are generic enough to collide with other workloads.
Co-developed HireBuddy's GPT-4o-mini pre-screening engine, scoring candidates across 10 evidence-backed dimensions with anti-hallucination safeguards that override the model's self-reported matches using profile evidence. Built a LangGraph + MCP agentic workflow (KnowFlow) that automates InstructLab knowledge contributions, and replaced a quota-limited third-party LLM in TerrierTracker with a zero-marginal-cost vector-embedding semantic recommendation engine. I contribute upstream to the LLM inference stack itself: vLLM, llm-d, SGLang.
Undergraduate research assistant at BU's Center for Innovation in Social Science: conjoint survey and vignette design, and pre-analysis plan preparation, for dissertation research on financial-market responses to geopolitical uncertainty. I ran a pre-registered ablation on my own ML model and cut the graph component when the no-graph MLP won — 0.633 against 0.613 test macro-F1. In August 2026 I joined BU's H2X Lab (Human-to-Everything, College of Engineering) as an undergraduate research assistant, and this fall I am a Course Assistant for CS 132 (Linear Algebra / Geometric Algorithms).
BostonHacks 2025, DS+X 2025, and Best Use of AI among 70+ participants at the Red Hat–IBM OpenAccelerator Hackathon — where the solution was accepted for upstream evaluation into Red Hat's official InstructLab open-source project.
From Hong Kong to Boston — a CS student in BU's Kilachand Honors College who builds software for real people, not just demos.
Software engineering is the job I want and the work I already do: shipping production features across three internships, running an international nonprofit's entire ~300-endpoint platform as CTO and sole engineer, and maintaining TerrierTracker for 1,000+ students. What makes me unusual is the scale of the range around that core — 68 open-source pull requests across 18 organizations, 28 merged by core maintainers of projects like vLLM, llm-d, and Redis, from a benchmarking agent core to monitoring dashboards, every fix shipped with the regression test that proves it; root-causing a recurring production outage on a live university server; and undergraduate research at BU's Center for Innovation in Social Science.
Conversational AI agent that lets domain experts contribute knowledge to InstructLab models without touching Git or YAML: a LangGraph ReAct agent driving a 10-tool MCP server for taxonomy search, Q&A generation, YAML validation, quality checks, document chunking and GitHub PR creation. Won Best Use of AI among 70+ participants at the Red Hat–IBM OpenAccelerator Hackathon; accepted for upstream evaluation into Red Hat's InstructLab project.
A columnar massively parallel SQL engine modeled on Amazon Redshift: all 22 TPC-H queries match DuckDB at 6M rows (scale factor 1); a leader plus one worker process per slice, 8 column encodings, zone maps, sort keys with VACUUM, KEY/EVEN/ALL distribution, and a Postgres wire-protocol server that psql connects to.
BU Spark!'s project gallery and staff admin portal, built and deployed as sole author — including a template-aware parser that extracts structured project data from free-text Google Docs without an LLM, and multi-semester identity resolution that merges duplicate people and projects.
A red-team study of LLM résumé screeners: 8 attack classes (visible and white-text PDF prompt injection, hidden keyword stuffing, Unicode-tag, zero-width and homoglyph smuggling via PyRIT converters, delimiter/role spoofing, fabricated skills) against 4 screener designs over 36 synthetic candidate–job pairs on two locally hosted models (Llama 3.1 8B, Qwen 2.5 7B), with 95% bootstrap confidence intervals across 3,104 scored records.
An education platform for an HSCA affiliate, built as sole engineer across all 347 commits: 21 data models and 35 API routes covering Stripe subscriptions mapped to plans and roles, cohort booking hardened against seat races, and single sign-on from the HSCA Portal.
The official event site for Clash of the Titans — the world's largest in-person student film challenge, hosted by the American School of Paris, where nine international schools race to plan, shoot, and cut a short film in under 48 hours. Built the full site end-to-end with a cinematic, film-stock design system.
A subscription learning platform for IB Film educators and students — gated lessons, courses, and a blog with tiered Stripe subscriptions and role-based access (including automatic HSCA educator verification). Full-stack build with NextAuth, a Prisma/Neon Postgres backend, an admin content studio, and S3-backed uploads.
