What MCP actually changes about tool integration
The Model Context Protocol is often described as USB-C for AI tools. The more useful framing is that it moves the integration boundary.
I build the infrastructure that takes AI research into production.
I work on retrieval backends, agentic tooling over the Model Context Protocol, and the deployment machinery underneath them: the unglamorous layer where a research prototype turns into something a customer can depend on. Before FTAI I spent two years in computer vision research at Washington University, which is where I learned what production has to survive.
Selected work
- 2025
Unified ML Feature Store
One feature definition serving both real-time inference and historical training queries, so the two never drift apart.
- 2025
Multimodal Video Q&A Engine
Ask a question about an uploaded video and get an answer with timestamped citations back to the footage it came from.
- 2024
GNSS/INS Sensor Fusion Filter
An error-state Kalman filter in Rust that fuses inertial and GNSS measurements, written to learn state estimation properly.
Research
- TerraBytes II @ ECCV 2026 ยท Best Paper Award
TerraDiT: Point-Conditioned Diffusion Transformer for Satellite Image Synthesis
Text-to-satellite image generation controlled by sparse point prompts instead of dense pixel-level layout maps.
- Preprint
Crossview Registered Multiview Pose Estimation
Recovering the pose of a set of street-level images inside a single aerial reference frame, plus the first public dataset for the task.
Experience
- 2025- Software Engineer FTAI Aviation
- 2025 Software Engineer Intern WashU Digital Innovation Accelerator
- 2024-2025 Software Engineer Intern WashU Multimodal AI Lab
Writing
Training/serving skew is a plumbing problem
The gap between how a feature is computed in training and how it is computed at inference is rarely a modelling failure. It is two codepaths.
Contact
The fastest route is email. I read everything, and I reply to anything that is not a template.