KAVISH SATHIA
Computer Science at NUS. I'm working toward the DWIM (Do What I Mean) compiler: building the tools that close the gap between what a human means and what a machine does.
SELECTED WORK
Star
A programming language that compiles to WebAssembly
A statically-typed programming language that compiles to WebAssembly, facilitating seamless cross-language library development. Write once, generate idiomatic APIs for Python, JavaScript, Rust, and Go. No more parallel SDK maintenance.
Sworn
The AI accountability framework
A framework for observing silent errors in AI agent executions at scale. Developers define behavioural contracts and a separate verifier evaluates deliverables against them. Features semantic, deterministic and NLI verifiers, plus a contract coverage metric to surface unmonitored behaviours.
Gauntlet
Adversarial fuzz-testing for AI agents
An autonomous adversarial agent that intercepts your AI agent's tool calls in real time, creatively manipulating results to discover security flaws like prompt injection, data exfiltration, and content poisoning. Uses a closed hypothesize-prove-store cycle with short-term and long-term memory circuits built entirely within Elasticsearch Agent Builder.
Vault
Privacy-first universal preference management
A system that eliminates the cold start problem by letting preferences follow users across apps. Raw text never leaves the device: client-side WebLLM generates semantic embeddings stored via pgvector, with game-theoretic anti-gaming and temporal decay.
Oz
An AI operating system built under 24 hours
An OS where software is an extendable primitive. Users generate, modify and publish applications via natural language, with Google Drive as the backing filesystem. Features an OzSDK mimicking Linux syscalls. Top 9 at Hack&Roll 2026.
