Experience

I founded Athlete Space and built it myself: an AI coach for endurance athletes that combines a multi-agent LLM system with established training-load science.

• A multi-agent architecture with separate planning, advising, evaluation and verification agents.
• A deterministic seven-stage planning pipeline, from season plan to daily calendar.
• A Python, FastAPI and PostgreSQL backend.

I ran applied AI at Wizard, a venture-backed AI company. My team built the agentic platform behind its search and conversation products, and I was responsible for the roadmap, hiring and architecture.

• Built the conversational engine, with agent memory, planning and tool use.
• Automated taxonomy generation and catalog enrichment, cutting manual maintenance by more than 90%.
• Connected live traffic to automatic evaluation, retraining and fine-tuning.
• Set the AI roadmap with product and go-to-market leadership, and ran hiring for applied science.

Four years as an applied scientist. I took ML systems for real-time decisions from prototype to production and kept them running.

• Modeled behavioral signals, with interpretability and live monitoring built in.
• Trained on real user behavior, not offline proxies.
• Worked with product, engineering and compliance to ship auditable models at lower infrastructure cost.

A research role during my Ph.D. I designed ML pipelines where latency, privacy and auditability were requirements from the start.

Ph.D. research on human pose estimation in the Computer Vision Laboratory, advised by Prof. Andreas Savakis. The work produced 20+ peer-reviewed papers, 500+ citations and one patent.

• Lead author on UniPose (CVPR 2020) and UniPose+ (IEEE TPAMI 2021).
• Mentored graduate researchers and managed sponsored research projects.

I built risk models on aviation-safety data to simulate drone collision scenarios, for the policy and operations staff who set the rules.

An engineering role before graduate school and the move into machine learning.