Projects
AI Document Research Agent
This project explores what it takes to build a genuinely autonomous AI agent — one that decides its own tool-use path in pursuit of a goal, rather than following a hard-coded workflow. Given a research question, the agent should be able to:
- Search a private document collection via retrieval-augmented generation (RAG)
- Decide whether it has enough information, or needs to search again with refined queries
- Synthesize an answer, citing which documents it drew from
- Save its findings for later retrieval
- Support natural follow-up questions within a session
The project is also a deliberate exercise in modern backend architecture — Spring Boot, PostgreSQL, vector search, containerization, and CI/CD — applied to a genuinely current problem domain (AI agent design) rather than a toy example.
Japa — Android Mantra Counter
A mantra-counting app for Android, built to solve a real problem that existing apps handle poorly: most japa counters require the screen to stay on and the user to tap the display for every repetition. For a practice that can run 1.5–2 hours a day, that’s impractical.
The key design decision: counting happens through the device’s volume buttons, so the screen can stay off for the entire session. Getting this right turned out to be the hardest engineering problem in the project — capturing hardware key events while the screen is off requires a foreground service running on a background thread, and Android’s process lifecycle management (Doze mode, OEM battery optimization) actively works against you. Getting the threading model right so the service doesn’t get interrupted mid-session took real trial and error.
Other features:
- Customizable mantras, repetitions per round, and rounds per day
- Audio chanting support — triggered by volume button or set to repeat continuously, with adjustable playback speed
- Dual-script text display: Roman transliteration and Devanagari
- Additional prayers screen for reference
- Automated build pipeline via GitHub Actions
Status: Alpha testing
GPU Covert Channel Security Research
Graduate-level security research (2015) demonstrating a novel attack vector: malware that hides its activity entirely within a GPU, out of reach of CPU-bound security software.
The attack: A Windows kernel mode driver written in C exploits Direct Memory Access (DMA) to locate and monitor the system keystroke buffer without CPU involvement, rendering it invisible to contemporary anti-malware software. Captured keystrokes are encrypted and exfiltrated to a remote system entirely from GPU memory — the CPU never touches the data after initial capture.
The defense: A companion detection tool leverages the PCI bus to inspect video card memory from a CPU-bound process, exposing GPU-resident malware activity. Performance characteristics of the detection approach were analyzed and mitigation strategies identified.
This project involved kernel mode driver development (KMDF), CUDA GPU programming, DMA exploitation, and PCI bus inspection — a stack most software developers never touch. The research demonstrated both the attack vector and a viable detection countermeasure.
The full project report and presentation are available in the repository.
Personal AWS Project
In active development — details coming soon
A personal project I’m using to build hands-on experience with the cloud-native patterns I want to bring into my professional work. Current stack includes:
- AWS — Lambda, S3, CloudFormation, serverless Valkey for caching
- Docker — containerized PostgreSQL with persistent volume storage
- AI-assisted development — using Claude for scaffolding, troubleshooting, and design tradeoffs in a real engineering context
This is where I’m closing the gap between the enterprise Java background I’ve built over 25 years and the modern distributed, cloud-native architecture the industry has moved toward.
IBM OnDemand Migration
Federal system — code not publicly available
One of the more challenging projects of my career, worth describing even without a code link.
At Carelon Federal Services, document storage for a federal platform serving 3M+ users was running on IBM OnDemand, with annual licensing costs in the millions of dollars. I volunteered to lead the migration to an in-house Vault system — a project nobody else wanted to own, because once committed there was no rollback path. A mistake meant permanent loss of critical federal records.
I designed the migration approach, built the integration layer in Java using REST calls to the Vault API, validated data integrity at each stage, and executed the migration without losing a single record. The result: several million dollars in annual licensing costs eliminated.
The technical work was a plain Java library consumed by both legacy Struts applications and a Spring Batch application — nothing exotic, but the stakes made it memorable.