> ryan_witzman
Open to Work
// shipped & open-source

Technical Registry

Open-source models, datasets, and specialized architectures — built end-to-end, trained on real GPUs, deployed to real workloads.

Computer Vision Fall 2025

GolfCart YOLO11s

A lightweight object detection model optimized for identifying golf carts in aerial drone imagery. Custom-annotated dataset and fine-tuned architecture balancing edge-device inference speed with high precision.

mAP@500.93
Precision0.978
Model Size22 MB
Latency14 ms
build_log — how it was made
  • Captured and hand-annotated a custom aerial dataset from drone footage — bounding boxes drawn frame-by-frame for varied altitudes and angles.
  • Chose YOLO11s over larger variants to hit a 22 MB footprint suitable for on-drone / edge inference.
  • Tuned augmentation (mosaic, HSV shift) to generalize across grass, cart paths, and shadows.
from ultralytics import YOLO

# Load fine-tuned model
model = YOLO('best-3.pt')

# Run inference on aerial drone footage
results = model('test_images/golfcart_drone_01.jpg')
results[0].show()
🤗 View on HuggingFace
NLP / Generation Fall 2025

HaikuGPT-17M

A 17-million-parameter decoder-only transformer trained from scratch on a curated haiku corpus. Despite its size, it learns strict 5-7-5 structure and semantic coherence without prompting.

Parameters17M
Ctx Window64 tok
ArchGPT-2
Eval Loss1.84
build_log — how it was made
  • Trained from random init — no pretrained weights — to study what structure a tiny transformer can learn from a narrow domain.
  • Curated and cleaned a haiku corpus; tokenizer and 64-token context sized to the form itself.
  • The model internalizes 5-7-5 syllable structure purely from data — no constrained decoding or templates.
from transformers import pipeline

pipe = pipeline("text-generation", model="rwitz/haiku-17m")
print(pipe("The summer sun", max_length=20))
🤗 View on HuggingFace
CV / OCR 2025

Pinball Score YOLOv11s

A lightweight detection model trained on A100 GPUs to identify and segment digital scores on vintage pinball machines — engineered to handle extreme glare, reflections, and low-light arcade conditions.

mAP@500.99
Recall0.989
Data1.1k img
Epochs100
build_log — how it was made
  • Labeled 1.1k frames from real arcade footage — the hard part was glare and mirror reflections off glass-covered displays.
  • Trained 100 epochs on A100s; near-saturated metrics (0.99 mAP@50) on held-out frames.
  • Stream-mode inference crops score regions from live video for downstream digit OCR.
from ultralytics import YOLO

model = YOLO('pinball-best.pt')
results = model('arcade_clip_04.mp4', stream=True)

for r in results:
r.save_crop('output/')
🤗 View on HuggingFace