Hi! I'm Ben Steinig 👋

a software engineer focused on web and conversational AI development. I like exploring how AI and LLMs are changing the way we build and use software, and enjoy turning those ideas into real, usable projects.

Right now, I’m at Microsoft in San Francisco 🌉 working on creating a more connected workplace for millions of users. If you’d like to chat or collaborate, feel free to reach out.

GitHub LinkedIn Blog

Projects

P-CLAHE: Parallel Contrast Limited Adaptive Histogram EqualizationFeatured

A CUDA based parallelization of the CLAHE algorithm in order to perform real-time underwater color correction.

View details
CUDAC++Python
CodePaper

P-CLAHE: Parallel Contrast Limited Adaptive Histogram EqualizationFeatured

A CUDA based parallelization of the CLAHE algorithm in order to perform real-time underwater color correction.

  • Built a real-time CUDA implementation of CLAHE (P-CLAHE) to enhance underwater robotic vision, enabling autonomous submarine systems to run image preprocessing on limited GPU resources.
  • Improved performance by applying shared memory caching, loop unrolling, and warp-optimized parallelization, cutting execution time and reducing memory overhead at scale.
  • Delivered a 3–5x speedup over CPU and unoptimized implementations, validating scalability across diverse image sizes and demonstrating reliability for deployment in constrained embedded environments.
CodePaper

EmojiComplete: Inferring Emojis Based on Sentence Meaning and Sentiment

a DistilBERT-based model that predicts contextually relevant emojis from sentence meaning, achieving state-of-the-art accuracy for real-world messaging applications.

View details
PythonPyTorchSLURM
CodePaper

EmojiComplete: Inferring Emojis Based on Sentence Meaning and Sentiment

a DistilBERT-based model that predicts contextually relevant emojis from sentence meaning, achieving state-of-the-art accuracy for real-world messaging applications.

  • Developed EmojiComplete, an NLP model leveraging DistilBERT to predict emojis from full-sentence meaning and sentiment, improving contextual relevance beyond word-level autocorrect

  • Built and curated a modern 150k-tweet dataset, introducing emoji clustering to reduce label noise and boost prediction accuracy across diverse demographics

  • Achieved 39% top-1 and 78% top-5 accuracy, more than doubling prior benchmarks (DeepMoji), validating performance for real-world messaging and autocomplete applications

CodePaper

Underwater Image Enhancement and Detection

A novel image processing pipeline for image enhancement and detection designed to improve performance in murky underwater environments. Produced as an EECS 442: Computer Vision capstone project.

View details
Computer VisionPythonRoboticsOpenCV
CodePaper

Underwater Image Enhancement and Detection

A novel image processing pipeline for image enhancement and detection designed to improve performance in murky underwater environments. Produced as an EECS 442: Computer Vision capstone project.

  • Built a computer vision pipeline for underwater object detection, combining white balancing, CLAHE, Canny edge detection, and custom color filtering to enhance image clarity and identify navigation waypoints.

  • Designed and implemented algorithms from scratch in Python/NumPy, benchmarking against OpenCV to evaluate both visual accuracy and runtime performance.

  • Applied the system to a dataset of 2,800+ underwater images collected in robotics lab conditions, demonstrating effectiveness for autonomous submarine navigation and environmental monitoring.

CodePaper

Let's build something great

Have a project, role, or idea in mind? I'm happy to chat.

Email me