Summer 2026 · Miami, FL
Worked as a Machine Learning Intern at Detect, a startup building asset-management software for electric utilities. Owned the damper-defect project end to end, from dataset creation to a validated detection and defect-flagging pipeline for aerial power-line inspection imagery, then moved on to R&D for component re-identification.
What I worked on
- A damper detector for drone inspection imagery, reaching 0.99 F1 on a held-out test set, by building and annotating a 5,250+ box dataset, training a YOLO11s model with tiled high-resolution inference, and writing a custom containment-based NMS that raised precision from 0.87 to 0.99
- A rotation-invariant geometric representation of power-line structures (symmetry axis, per-conductor span lines) layered on detector outputs, used to flag slid and missing dampers that standard classifiers can't see and surfacing 75+ field-confirmed defects across ~1,000 inspection frames
- Evaluation and handoff for the full pipeline: an extrapolation study on unseen inspections and new hardware types, and a 530-frame held-out set screened for zero leakage, pre-annotated by the frozen detection, geometry, and defect stack for human review
- R&D on a DINOv3 component re-identification pipeline, re-engineering the embedding readout strategy to raise cross-viewpoint matching by up to 40% (0.63 → 0.88 AUC), validated on a 2,500+ pair benchmarking harness I built