// projects
Selected work
Four projects that cover the range of what I build: agentic AI systems, applied deep learning, full-stack web, and production mobile. Each one is described as it actually was built, limitations included.
Rosa — Multi-Agent AI Health Platform
April 2025
A multi-agent system for physical and mental health support, built by a 3-person team against a 25,000-user MVP target.
- A task-decomposition and routing layer takes a user query, breaks it into sub-tasks, and dispatches each to a specialized agent — one for physical health, one for mental health.
- Each agent carries its own tool set, including retrieval against a vector database holding session-wise memory, so answers are grounded through retrieval-augmented generation rather than pure generation.
- A response aggregator collects the specialized agents’ outputs, resolves overlap between them, and produces a single coherent answer.
- I built an evaluation process to score agent outputs against expected behavior, spot failure patterns, and iterate on routing logic, retrieval parameters, and prompts. That evaluation was largely manual — a maintained set of representative queries reviewed by hand.
What I’d do differently today
- 1.Use LangGraph rather than vanilla LangChain, so agent handoffs are an explicit state graph instead of chained calls.
- 2.Build automated evaluation from day one — instrument every agent interaction and score outputs continuously instead of reviewing them by hand.
- 3.Add real observability. Structured logging and tracing across agent steps, because working out which agent or retrieval step produced a bad answer took longer than fixing it.
- Python
- LangChain
- RAG
- Vector DB
- Agentic AI
Knee Osteoarthritis Detection Using Bone Distances
PublicationIEEE/ACM CHASE 2025 · New York, June 2025
Applied deep learning on medical imaging: predicting osteoarthritis severity from geometric bone-distance measurements.
- A MATLAB–PyTorch supervised deep learning pipeline analyzing 160+ MRI scans to predict osteoarthritis severity from geometric bone-distance measurements.
- Achieved 76% accuracy and an AUC of 0.78.
- First-authored with Tarun Ramapuram and Dr. Juan Shan; presented at IEEE/ACM CHASE 2025 in New York City, June 2025.
- Framed accurately: this is applied deep learning on medical images driven by geometric measurement — model design, training, and evaluation — not detection, tracking, or segmentation work.
- PyTorch
- MATLAB
- Deep Learning
- Medical Imaging
Interactive Novel Web Application
February 2025 – May 2025
A full-stack app generating personalized storylines in real time from the reader’s choices — front end, backend, database, and cloud, owned end to end.
- Generates personalized, real-time storylines from user choices, with real-time narration and emotion-driven thematic elements.
- Integrated the OpenAI API for real-time content generation inside a production-style REST service.
- Deployed on Amazon EC2 (Linux), with Amazon RDS for backend data services and CloudFront caching image delivery.
- Built session tracking and secured backend endpoints, then profiled and optimized request flows — cutting API call volume by 30% while holding reliability steady.
- React
- TypeScript
- Node.js
- OpenAI API
- MongoDB
- AWS
Green2Gold Mobile App
Professional workJune 2025 – Present
An end-to-end mobile app for an environmental education organization: Flutter client, Python REST API, PostgreSQL data layer.
- Built and maintain a Python REST API backend paired with a Flutter mobile client, supporting content delivery and subscription workflows.
- Designed a PostgreSQL data layer with clean, reusable endpoints covering consultant inquiries, event sponsorship forms, and media assets.
- I had no mobile development experience going in. I learned Flutter on the job and shipped production features within weeks.
- Flutter
- Dart
- Python
- REST API
- PostgreSQL
Internal app — the organization has not published it yet.