AI for the Visually Impaired: A Constraint-First Assistive Vision Report
A constraint-first AI assistive-vision project for blind/low-vision users that trains models under real deployment limits (offline, low latency/memory, noisy "blind photography", safety-critical operation) rather than optimizing clean-data accuracy first. This approach yields better calibration and smoother degradation than scale-first distillation, and proposes a modular offline-first system spec (OCR, object/currency recognition, motion cues) aligned with EU regulatory requirements.
See Project