Blog
Stories, updates,
and engineering.
Behind the scenes of building Gestura — model training, product decisions, and what's coming next.
How we built Gestura from scratch in 4 phases
From a blank Python file to a live cloud-deployed ASL translator — here's the full story of how Gestura was built, what broke, and what we learned.
Gestura origin story
Training an LSTM model on 87,000 ASL images — what we learned
We processed 87,000 hand gesture images through MediaPipe, extracted 63 landmark coordinates per image, and trained a custom LSTM. Here's exactly how we did it and what 98.2% accuracy actually means.
Why we use MediaPipe landmarks instead of raw images
Most ASL projects train on raw images. We don't. Here's why extracting 21 hand landmarks first made our model faster, smaller, and more accurate.
Getting real-time inference under 200ms with FastAPI
Real-time means nothing if your API is slow. Here's how we optimized our FastAPI + PyTorch inference pipeline to consistently hit sub-200ms response times.
Gestura roadmap — what's coming next
Word-level detection, mobile support, multi-language sign languages, and a public API. Here's everything we're building next and why.
Why accessibility tech matters more than ever
Over 70 million people use sign language as their primary language. Here's why we believe AI-powered accessibility tools are one of the most important things we can build right now.