Machine Learning Engineer @ Owl AI
I build real-time computer-vision systems that run live on broadcast television — ball tracking, trick metrics, and instant officiating delivered to national audiences in seconds.
Third engineer on the founding team at Owl AI — live on X Games broadcasts and Fox Sports MLP pickleball coverage (PickleballTV, FOX & YouTube).
Real-time systems running on live sports broadcasts. A wrong call goes out to a national audience — so reliability is the whole game.
Seven broadcast cameras feed the system at 240 frames per second over an SRT link.
GStreamer ingests every stream with GPU-optimized reads — nothing dropped on the way in.
A Triton-hosted model zoo — RF-DETR, TrackNetV5, SAM — tracks the ball and pinpoints the bounce.
The verdict — in or out, and how close to the line — is rendered as a broadcast graphic in Godot.
The call hits the national broadcast within 5–10 seconds of the bounce.
Definitive in/out calls for Major League Pickleball on PickleballTV, FOX & YouTube — delivered within 5–10 seconds of the bounce.
Autonomous skater tracking that detects trick start/end for height, airtime and trick type — feeding a live judging console at X Games Sacramento, Chiba & New Orleans.
Athlete tracking and total-rotation computation with GStreamer broadcast ingestion — tricks land on the judging console the instant they finish.
Live dunk-contest metrics — peak height above the ground, dunk power and jump distance — computed in real time through heavy on-court clutter.
Automatic rally segmentation so coaches get only the rallies — cutting annotation from ~10 hours to ~90 minutes and halving monthly labor cost.
English race commentary re-voiced in Japanese — ASR + diarization, staged LLM translation, per-commentator TTS voices, timing-fit to stay in sync with the broadcast.
Sports is the current chapter — the same engineering shipped enterprise AI products and peer-reviewed research first.
Co-architected a Microsoft-certified Teams assistant that answers from large document corpora with grounded, source-backed responses — ingestion guardrails, hybrid + vector retrieval with re-ranking, and Redis-cached Node.js services.
Segmentation of radiation damage across four microscopy modalities, and tensile-property prediction from sparse data (RF/XGBoost, Pearson r > 0.98 with GAN/SMOGN augmentation) — co-authored in two Frontiers journals.
Multi-armed-bandit traffic allocation with group-sequential significance thresholds — decisions 46% faster, with BigQuery-to-Tableau analytics reporting in near real time.