Building on-device AI companions while investigating whether weak models can catch reward hacking.
AI SystemsAI SafetyEdge ComputingPrivacy
Kaushik is a London-based systems engineer and founder building a privacy-first AI companion centered on on-device intelligence and local models. Alongside product work, Kaushik pursues open research into AI alignment questions, such as whether weak language model monitors can distinguish between reward hacking and failing agents.
What the sources establish. Gold means missing.
From Edge Devices to AI Supervision
Both efforts focus on containment and resilience, ensuring models remain secure locally and honest when evaluated.
In their wordssystems engineer, building @poppys_social ,prev @Web3Chennai, @ProofOfWitness , @b2fdotfun , @PinaivuAi
The portrait and roast are playful interpretations of public sources. The factual claims were checked against the saved sources; they are not independent verification.
Working on
ai
Craft
research
Building style
Not yet known
Founding role
solo
Categories are model inferences. “Not yet known” means the sources do not establish a category.
Kaushik is a systems engineer and founder based in London.
Open research: can a weak LLM monitor tell reward hacking apart from an agent that's just failing? Both leave the score above reality.
Collaborators and critique welcome: @bshlgrs @RyanGreenblatt @bobabowen
https://t.co/owDZRbhHc4
my thought about the question: what if we simply can’t stop powerful AI capabilities from spreading?
instead of treating AI safety only as “slow capabilities down,” the idea is to deliberately accelerate technologies that make society harder to attack.
take cybersecurity.
if
I’m a founder in London 🇬🇧
building a privacy-first AI companion around local models, on-device intelligence, and the idea that your AI should know you without owning your data.
would love to meet more people building weird, ambitious things in AI.
