ayush sahu

Building AI trust infrastructure for the AI-native internet.
CTO at DetectifAI, focused on real-time audio deepfake detection.

[ayush sahu @ayus_detectifAI]

@SoundPapers this gets really interesting once the audio goes through phone compression and replay. clean generated clips are the easy part now.

[ayush sahu @ayus_detectifAI]

@tchsignal good distinction. an incapable robot can look safe by accident, while the policy behind it is still making the wrong call.

[ayush sahu @ayus_detectifAI]

@0x0SojalSec this is such a weird failure pattern. refuses the knife every time, then happily puts compressed air on a hot stove. refusal rate alone clearly isn't safety.

[ayush sahu @ayus_detectifAI]

@scholarpulse 97% attempted sounds terrifying, but completion and intervention matter too. the benchmark gets much more useful when those are separate numbers.

[ayush sahu @ayus_detectifAI]

@sreexts @aelluswamy liability will force better safety faster than another leaderboard. when a robot hurts someone, 'the model passed our eval' won't be much of a defence.

[ayush sahu @ayus_detectifAI]

@EntelligenceAI per-step routing is where this gets interesting. planning, tool use and final synthesis probably shouldn't all go to the same model.

[ayush sahu @ayus_detectifAI]

@omiossec_med routing looks easy until cost and quality disagree. saving 30% means nothing if reliability quietly changes from one request to the next.

[ayush sahu @ayus_detectifAI]

@AIMarketMind the real moat might be identity + memory, not the model. if users can carry both to another agent, the platform suddenly looks a lot less sticky.

[ayush sahu @ayus_detectifAI]

@hackinarticles poisoned retrieval data still feels under-tested. people secure the prompt, then let the agent trust whatever lands in its context.

[ayush sahu @ayus_detectifAI]

@cao_lab @din_lol_ exactly. the scary agent isn't always breaking the rules - sometimes it's using permissions we gave it too broadly.

[ayush sahu @ayus_detectifAI]

@docsmsft this handoff can be really useful if the rep gets a clean result, not another wall of transcript. consent captured, PIN verified, back to the human.

[ayush sahu @ayus_detectifAI]

@_memCode the underrated part is deciding what *not* to remember. otherwise the agent gets smarter for a week and noisier forever.

[ayush sahu @ayus_detectifAI]

@_jaydeepkarale finance agents are where 'mostly correct' stops being good enough. a right answer from stale data can be more dangerous than an obvious failure.

[ayush sahu @ayus_detectifAI]

@AverageAiBro @DanKornas this feels much closer to how memory should work. embeddings are useful, but without the who/when/why they turn into a very confident junk drawer.

[ayush sahu @ayus_detectifAI]

@beryIxo Would you run both versions of the eval: one with memory reset and one with the real accumulated state? The gap between them might be the most useful metric.

[ayush sahu @ayus_detectifAI]

@imsaqlain22 How do you handle external tools that change underneath the trace? A replay seed helps, but API versions and live data can still make the same failure impossible to reproduce.

[ayush sahu @ayus_detectifAI]

@Coffee_and_NLP @zohaibahmed @resembleai The channel seems like the real stress test now. Are detectors being evaluated after phone compression and re-recording, or mostly on clean generated audio?

[ayush sahu @ayus_detectifAI]

@vigram_void Runtime shielding makes sense, but how well does the constraint transfer to unseen objects and tasks? A safety layer that only works inside the benchmark would be easy to overtrust.

[ayush sahu @ayus_detectifAI]

@jkq6000 Do you think barge-in should adapt to intent? Interrupting a confirmation sounds useful, but interrupting while the agent reads a critical number could be dangerous.

[ayush sahu @ayus_detectifAI]

@SoundPapers Latency gets measured a lot, but interruption recovery matters just as much. Do they report how often the agent resumes the right intent after a barge-in?