Antony Ranjith

A system that estimates, while the call is still going, whether the speaker is a person or a voice agent. I work on the speech-model pipeline, on evaluation against both real and generated speech, and on code review across the codebase.

The project is private, so names and dataset figures stay off this page. I am glad to walk through the approach and the eval design in an interview.

• Python
• PyTorch
• self-supervised speech models
• evals

OpenVox

Private, on-device voice dictation for macOS. Side project.

Hold a key, speak, and the text lands at your cursor in any app. Two streaming modes: Moonshine Streaming transcribes when you release the key, and NVIDIA Nemotron Speech Streaming 0.6B types while you speak. The models run locally on Apple silicon, so audio never leaves the machine. The app downloads each model the first time you use it.

• Swift
• Python
• ONNX
• MLX
• streaming ASR
• on-device inference

Download for macOSSource

A compact System 1 decision model. The hypothesis: rule-conditioned computation and counterfactual training help a small network generalize to new decision tasks better than a standard transformer classifier. Four architectures, from an Ettin-17M baseline to an instruction-conditioned hypernetwork with permutation-equivariant scoring, are compared under one pre-registered protocol with fixed seeds and equal budgets.

Data pipeline, counterfactual generator, and frozen evaluation splits are done. Model training has not started, so there are no results yet.

• Python
• PyTorch
• hypernetworks
• counterfactual learning
• evals

I write the architecture and split the work into independent tracks that coding agents (Claude Code, Codex) can run in parallel. The agents work from standards docs I wrote for them, and every PR goes through automated review → fix → re-review loops before a human signs off. When the tooling falls short, I build my own.

• Core protocol (Move): I owned the architecture of the clearinghouse, positions, liquidations, funding, TP/SL and scheduled orders, market modes (active, reduce-only, MM-only, settlement), role-based administration, and the Pyth oracle integration.
• Off-chain stack: Pyth price keepers with batched on-chain updates, a scheduled executor, a funding bot, and market-making bots across six markets.
• SDK: the public TypeScript client @xyra-trade/perps, with REST and WebSocket transports and order builders for limit, market, stop, and take-profit orders.
• Now: leading the build of a next-generation, high-performance exchange engine in Rust.

• Move
• Rust
• TypeScript
• Aptos
• Pyth
• risk engines

• Built on Aptos fee-payer transactions (AIP-39): dApp registration, user and function allowlists, sponsored submits, and fee-payer pool management.
• Designed and operated the production service (TypeScript, Prisma, Kubernetes) and shipped @kanalabs/paymaster-sdk.

• TypeScript
• Kubernetes
• account abstraction
• AIP-39