projects

03
  1. 01

    Grocery Item Counting with Vision-Language Models

    Counting the items in a top-down photo of a grocery cart, toward checkout without barcodes or RFID. With two collaborators, I annotated a 612-image benchmark and compared zero-shot Gemini, a LoRA-fine-tuned Qwen3-VL-2B, and a YOLOv8 detector.

    fine-tuning cut Qwen3-VL-2B’s counting error from 2.10 to 1.52 MAE

    Python, PyTorch, LoRA, Qwen3 VLM, YOLOv8, Google Gemini

    2026

  2. 02

    Predict NFL Offense

    Reads a formation and game situation and predicts each receiver’s route and whether the play is a pass or a run — or flip it around: design your own play and see the likely outcome. Every route on the field is drawn from real 2022 NFL player tracking.

    76.4% pass/run accuracy on unseen games (vs. 60% baseline) · SHAP explanations for every call

    Python, LightGBM, SHAP

    2026

  3. 03

    Impact Tackling

    Tackling scores for every defensive player tracked in weeks 1–9 of the 2022 season, built on NFL Big Data Bowl tracking data, then compared across positions and against Pro Bowlers.

    R

    2024

more on github