Web-hosted demo (Hugging Face): https://huggingface.co/spaces/0saker/temseg-demo

GPU-backed deployment (Modal): https://malhotrasimar009--temseg-fastapi-app.modal.run

A quick note on performance: the Hugging Face demo runs on shared hosted infrastructure and can take ~30–50 seconds for a full segmentation job. The Modal deployment uses GPU acceleration and is much closer to the desktop application’s performance (typically ~5 seconds on representative images), but is a lot more fragile atm

Technical spec writeup: https://simar-malhotra09.github.io/writing/documents/draft1.html

TEMseg, a desktop application our research lab built to automate nanoparticle analysis from transmission electron microscopy (TEM) images:

I led the initial proposal, UX design, and end-to-end technical implementation. A key part of the project was working closely with postdocs and PhD researchers in materials science to validate both the technical approach and the product-user fit.

The main challenge was designing a system that could run two state-of-the-art computer vision models in a way that still felt interactive for researchers. We achieved sub-second refinement operations by caching model embeddings and optimizing inference with ONNX Runtime. The application extracts per-particle morphology statistics calibrated to physical units and fits size distributions against reference models.