Serginusa

Really impressive stack — especially the quantization workflow with TensorRT/INT8 on Jetsons. We've been dealing with similar tradeoffs (speed vs segmentation accuracy) in other domain:

Curious — how many labeled fish images did you need before the quantized models stopped falling apart in production?

(Also, for anyone tracking W26, we've got OctaPulse on our prediction market: ingene.win/?utm_source=hn_comment&utm_medium=social&utm_campaign=mar2026)