DEMOS

Two bits, one changed fact — following a semantically targeted bit-flip search through a speech-recognition model, from audio to altered transcript. Interactive explainer.

Bit-level fault analysis and semantic steering in a quantized image captioner — interactive explainer.

Protecting AI training from corrupted human feedback — product demo.

Anchor-calibrated trust estimation — recovering trusted vs. corrupted sources when the majority is wrong. Interactive explainer.

Commercializing COBRA — product framing, learnings, and a possible repositioning of the research as a trust layer.