An event-driven, serverless security pipeline on AWS that scans every pull request for vulnerabilities and posts findings back as inline comments. Engineered for reliability — SQS retries, a dead-letter queue, CloudWatch alarms, and one-command infrastructure-as-code.

A multi-agent RAG system (answer + search agents) that took first place, paired with a headless browser-automation pipeline and real-time instrumentation for accuracy, latency, and end-to-end response time.

Say "take me to table 6" and a real TurtleBot 4 plans and drives the route. An LLM planner turns natural language into tool calls, a semantic map resolves the destination, and ROS 2 Nav2 with AMCL does the driving, recovering on its own from clock drift and localization collapse. I built the LLM planner and its tool-calling loop (retries with backoff, mid-loop cancellation, health aggregation, history trimming), owned CI/CD, QA, and end-to-end testing across a 114-test suite, and mapped the lab with SLAM. Five-person team project, CS5335 Robotics at Northeastern.

A full-stack RAG system for Massachusetts tenant law, built end to end: custom web scrapers and a 967-chunk corpus, a retrieval pipeline comparing eight strategies — from vector and BM25 to hybrid fusion, cross-encoder reranking, parent-child expansion, multi-query, and auto-merge — over a ChromaDB vector store, and a Next.js/React chat UI with inline citation cards on a streaming FastAPI backend. I designed a bias-controlled evaluation harness (nine model-by-retriever configurations, LLM-as-judge from a separate model family) that measures faithfulness, relevancy, and retrieval quality, and showed retrieval grounding lifted answer faithfulness roughly sevenfold over a no-retrieval baseline under a strict separate-family judge.

An open-source, cross-family LLM-as-judge for RAG evaluation — it scores faithfulness, relevancy, correctness, and retrieval metrics while keeping self-evaluation bias out. Extracted and generalized from my Renters eval harness after a same-family judge (GPT-4o grading GPT-4o) flattered a no-retrieval baseline to 0.64; switching to a cross-family judge (Claude Sonnet 4) exposed the honest 0.10 — turning a ~1.4× "uplift" into the real ~7×. Bring your own outputs: it defaults to cross-family judging and warns when the judge shares a family with the generator. ~570 LOC; deps are just openai + python-dotenv via OpenRouter.

Product & technical lead on a modular research platform (14 spec'd modules) comparing multi-agent code-review approaches — owning research direction, system design, data extraction, and visualization.

A public-facing Reconciliation Action Plan (RAP) data portal for Indigenomics AI — report → confirm → coverage across company, supplier, and institute personas. Next.js on AWS Lambda with a single-table DynamoDB design and a Bedrock + Textract extraction pipeline, deployed across a Canada/US data-residency split.

AI language learning across 20+ languages — spaced repetition, interactive stories, games, and vocabulary from any video or image.

AI-native presentations — hassle-free, clicker-free, voice-controlled, and context-aware presentation flow.