Portfolio

I work independently on consumer mobile apps and build the whole thing myself, from the idea and the interface down to the shipping pipeline.

I work independently on consumer mobile apps and build the whole thing myself, from the idea and the interface down to the shipping pipeline.

Researched human pose estimation in complex, real-world video where occlusion, motion blur and crowding break the usual assumptions.

Wrote a state-of-the-art review of pose-estimation methods and proposed a step-by-step method to improve regression accuracy.

Researched human pose estimation in complex, real-world video where occlusion, motion blur and crowding break the usual assumptions.

Wrote a state-of-the-art review of pose-estimation methods and proposed a step-by-step method to improve regression accuracy.

Worked on a badminton match-analysis system: court detection, shuttlecock tracking, and the pipeline tying them together.

Introduced a court-detection method based on semantic segmentation that reached 99.08% test mIoU, and improved shuttlecock-tracking accuracy by 5%.

Made the whole system run four times faster, mostly by threading the heavy work.

Worked on a badminton match-analysis system: court detection, shuttlecock tracking, and the pipeline tying them together.

Introduced a court-detection method based on semantic segmentation that reached 99.08% test mIoU, and improved shuttlecock-tracking accuracy by 5%.

Made the whole system run four times faster, mostly by threading the heavy work.

Worked on MetaSport, a metaverse attempt at bringing sports together in one digital place.

Learned the Unity engine in a week and built a demo sports game driven by reinforcement learning.

Worked in a team on the 3D-world infrastructure and designed over 20 features on Unity.

Named top-performing intern of the cohort for the consistency and quality of the results.

Worked on MetaSport, a metaverse attempt at bringing sports together in one digital place.

Learned the Unity engine in a week and built a demo sports game driven by reinforcement learning.

Worked in a team on the 3D-world infrastructure and designed over 20 features on Unity.

Named top-performing intern of the cohort for the consistency and quality of the results.

A Python/Tkinter desktop app that transcribes the audio of a YouTube video so an agent can summarize it and push the summary straight to Notion.

Made for getting a quick summary of a video instead of taking notes.

One of my first projects built with Claude Code, to learn agentic coding and context engineering.

A Python/Tkinter desktop app that transcribes the audio of a YouTube video so an agent can summarize it and push the summary straight to Notion.

Made for getting a quick summary of a video instead of taking notes.

One of my first projects built with Claude Code, to learn agentic coding and context engineering.

A React front end over two Flask microservices — one for translation and language detection, one for text-to-speech.

Deployed on Azure with Kubernetes and Helm; infrastructure provisioned with Terraform.

Observability throughout: metrics in Prometheus + Grafana, tracing with SigNoz, logs in Datadog.

A React front end over two Flask microservices — one for translation and language detection, one for text-to-speech.

Deployed on Azure with Kubernetes and Helm; infrastructure provisioned with Terraform.

Observability throughout: metrics in Prometheus + Grafana, tracing with SigNoz, logs in Datadog.

A computer-vision model that classifies facial expressions from live video, wrapped in a Chrome extension so it runs over video calls and YouTube.

Reached 63% validation accuracy on the FER2013 dataset.

A computer-vision model that classifies facial expressions from live video, wrapped in a Chrome extension so it runs over video calls and YouTube.

Reached 63% validation accuracy on the FER2013 dataset.

Pure-Python contrast, filtering, binarization and morphology, with an interactive Tkinter GUI to preview images and their statistics.

Made the filter size configurable and brought filtering down from O(n³) to O(n² log n).

Pure-Python contrast, filtering, binarization and morphology, with an interactive Tkinter GUI to preview images and their statistics.

Made the filter size configurable and brought filtering down from O(n³) to O(n² log n).

Ran Python lab sessions for two first-year classes (French and international tracks) and C lab sessions for a first-year class.

Supervised and marked end-of-module projects.

Ran Python lab sessions for two first-year classes (French and international tracks) and C lab sessions for a first-year class.

Supervised and marked end-of-module projects.

Ran a LinkedIn page focused on helping students through their graduation projects, and gave free consultations to guide them.

Ran a LinkedIn page focused on helping students through their graduation projects, and gave free consultations to guide them.

Second author on A Deep Learning-Based Framework for Racket Sports Court Registration, published in Artificial Intelligence Applications and Innovations (AIAI), Springer, 2024.

The work researched badminton court registration with image-segmentation models.

Second author on A Deep Learning-Based Framework for Racket Sports Court Registration, published in Artificial Intelligence Applications and Innovations (AIAI), Springer, 2024.

The work researched badminton court registration with image-segmentation models.

Picked up NVIDIA TAO and DeepStream in under eight hours and built a real-time video-inference pipeline that spots potholes in street footage.

Picked up NVIDIA TAO and DeepStream in under eight hours and built a real-time video-inference pipeline that spots potholes in street footage.