Colombian Sign Language (LSC) has far fewer digital resources than spoken Spanish, including the datasets a recognition model needs to exist in the first place. As a researcher at Universidad de Antioquia, I worked on that gap end to end: the corpus, the model, and the system that serves it.
Problem
There was no usable dataset for LSC recognition, and without one there could be no model. The research had to start below the model, with the data itself.
Solution
I worked the problem end to end: build and curate the corpus, train a recognition model on it, and wrap that model in a system people could actually use, not a notebook result.
Architecture
The public repository holds the application side: a Python backend built with FastAPI, exposing authentication, user management and sign-processing routes backed by MongoDB, with a React frontend, deployed on Kubernetes.
Engineering challenges
The corpus was the hard constraint the rest of the project depended on: building and curating a dataset of more than 1,500 videos before any model training could start.
My contribution
I worked on this as a researcher at Universidad de Antioquia across all three layers: corpus construction, model training (recurrent architectures: GRU and LSTM), and the FastAPI/React/Kubernetes system that serves the model.
Results
The model reached a 91.17% F1 score. The work is published as a 2024 journal article, Herramientas de aprendizaje automático con redes neuronales para el reconocimiento de Lengua de Señas Colombiana (Machine-learning tools with neural networks for Colombian Sign Language recognition), listed on ORCID.
Why it matters
Sign-language technology is a case where machine learning has a direct accessibility impact: every improvement in recognition quality moves toward communication tools for the deaf community in Colombia.