I build AI systems that people actually end up using.
Most of my work starts the same way: I go sit with the people doing the work by hand, figure out what they actually need, and then build something reusable instead of a one-off. Agentic LLM platforms, RAG, computer vision in production. The model is usually the easy part. Getting a team to trust it and adopt it is the job.
I'm finishing an MS in Information Systems Management at Carnegie Mellon (Cooper Fellow, Dec 2026), and I'm looking for Forward Deployed Engineer / Applied AI roles, where building and talking to customers are the same job. Open to relocating.
BNY · AI Hub (AI/ML Engineer Intern, 2026) Took a document-extraction pipeline that every team wired up differently and turned it into a self-service platform for building agents. Business teams describe a workflow, and the platform assembles, secures and ships the agent.
- 25+ agents across 6 production use cases, 10+ live deployments
- New use-case onboarding cut from a week to a day
- An estimated 155K analyst hours a year saved
- Found and fixed four silent failure modes nobody had flagged (access, context, tool scope, secrets), then added a review loop where every human correction becomes improvement data
Slamdunk.AI (Founding AI/ML Engineer, ~2.5 years) Joined an AI sports-coaching startup as an intern and left leading its recommendation team.
- Real-time athlete tracking (YOLOv8, MediaPipe, ONNX) at 96% detection accuracy, with cloud costs down 70%
- Wasn't happy with the coaching recommender I'd built, so I went and talked to tennis coaches, then rebuilt it: feedback latency went from 3 minutes to under 15 seconds
- Led a 3-person team that shipped the production MVP in 3 months
Research
- Now: keeping LLM agents from taking actions they shouldn't (OWASP LLM06, Excessive Agency). Graph-based access control on top of LLM Guard, red-teamed with AgentDojo.
- Published: graph neural networks for predicting which drugs cross the blood-brain barrier (IEEE 2023).
- Go to the user first. Ops teams, tennis coaches, a pharma analyst: the best spec I ever get comes from watching someone do the job.
- Don't trust "it works." A 0.935 mAP model that missed the ball in over 90% of real frames taught me to evaluate on the real thing, not the validation set.
- Build it so the next one is easier. A fix that only helps one use case is a patch. The goal is a platform.
Most of my professional work lives in private codebases. These are smaller things I build to learn and to scratch an itch, so they're a glimpse of how I think rather than a full picture.
| ⛳ SwingTrace | Open-source golf swing tracer. Tracks the club through motion blur in 93% of frames on unseen phone videos. Try it in Colab · Model |
| Autonomous SOC agent | Two LangGraph agents for security operations. One flags anomalous network traffic, the other drafts a recovery plan using a local LLM and a playbook knowledge base. |
| 🧠 BBBNet | Reproducible version of my IEEE paper: will this drug reach the brain? Graph neural network on 7K+ molecules. |
| golf-ball-tracking | Write-up of the "great validation score, useless in production" lesson, and the YOLO + SAM 2 tracker built after it. |
Stack I reach for: Python · TypeScript · LangGraph · MCP · RAG · PyTorch · YOLO · SAM 2 · FastAPI · Docker · AWS
📫 LinkedIn · [email protected] · Pittsburgh, PA