I'm an Information Systems master's student at Northeastern's Silicon Valley campus, working across UI development, LLM post-training and backend engineering. I've been a TA for Advanced Techniques with Large Language Models, for 2 semesters, and in my research I'm adapting multimodal models to geospatial reasoning with SFT, GRPO and Reinforcement Learning Reward Design.
Selected as 1/8 students university-wide for a fully funded exchange year in University of Florida, many of my landscape architecture projects happened here, from my capstone design in DeBary to a flood and coastline plan for St. Augustine. Doing the visualization and sea level analysis by hand took forever, so I taught myself to code to automate it. That's how I ended up with both a designer's eye and an engineer's habits. The St. Augustine pipeline below is that study rebuilt properly in code.
Now I build Python pipelines and tools that run unattended, check their own results and give the same output on every rerun: automated runs with retries and resume, regression tests in CI, versioned data and results. I'm looking for SDE roles in simulation and data tooling, and LLM engineering roles.
LLM failure mode discovery: flood reasoning from site maps
Research with Prof. Lai at Northeastern
Multimodal LLMs such as DeepSeek and Gemini miss almost every floodplain cell without mapped water (3/49 and 0/49), while a blind human gets 24 of 26. I built the 194-site dataset and the RL reward for this task, caught a reward hack along the way, and am now using post-training with SFT and GRPO on a reinforcement learning vision-language model.
The repo is private so future models can't find the answers. Contact me for details.
A breathing rehab game for people recovering from lung surgery at home: breathe into the microphone to jump and glide. I designed the trainer, the app and the illustrations, then built it in React and TypeScript, with unit and end-to-end tests gating every deploy. Two versions are playable: v1, the first release, and v2, with the redesigned app pages. React, TypeScript, Playwright, illustration |
How much of St. Augustine, Florida floods at 1 to 9 ft of sea level rise, rebuilt as a one-command Python pipeline on NOAA, Census and OpenStreetMap data, with byte-identical reruns, self-checks and CI on Linux, macOS and Windows. On the project page you can drag today against any level, cut a section through the town, tap a street to see when it floods, and see the year NOAA's scenarios bring each foot. Python, GeoPandas, pytest, GitHub Actions, JavaScript |
CloudBite Engine is a food delivery backend on Spring Boot, MyBatis, MySQL and Redis, with JWT auth and WebSocket order alerts. It includes 18 bilingual engineering notes where I question and verify the course material instead of copying it.
Patient management microservices is a patient management system split into five Spring Boot services that talk over gRPC and Kafka behind an API gateway, running locally on Docker and LocalStack.
Reach me on LinkedIn.