I'm Emma. I'm building an interpretable substrate for AI: a stack of geometric tensor systems meant to be the readable alternative to opaque neural-computer runtimes, in place before the opaque version locks in. The flagship is Yantra, an interpretable neural computer (a GPU-native operating system) written in my own language, Sutra, whose programs are vector symbolic architecture (VSA) operations compiled to tensor arithmetic.

I treat this as a venture, not a side project. Sutra, Yantra, and Loka are a coherent product stack with their own domains, a published paper, and working prototypes, and I'm deciding between founding the company around them and joining a safety lab. I build agent-first and high-output: most of what's here is driven by AI coding agents working inside my own scaffolding (cleanvibe), with the judgment to know when to take the wheel. Since March 2026 I have been shipping at a pace I did not know I had, across dozens of repositories.

• Sutra went public with a paper on arXiv (arXiv:2605.20919): an embedding-native language whose programs are VSA operations (bind, unbind, bundle) compiled to straight-line tensor ops and round-tripped back to verifiable source.
• Thermodynamic computing: early but measured experiments mapping Sutra's VSA operations onto a thermodynamic (thrml) sampling substrate. Associative memory, content-addressable retrieval, and bind/unbind all run as energy-based sampling on the hardware model, with measured recovery numbers rather than a sketch.
• Yantra went from nothing to a v0.0 kernel nucleus plus a hand-rolled no_std Rust bootloader that boots on bare metal in QEMU (PCI scan, GPU framebuffer write, long-mode transition).
• Loka trained role-aware transformers from scratch on Wikidata and shipped checkpoints to Hugging Face.

Bio, projects, research, and interactive tools all live there. The project subdomains below are the current homes — older standalone domains (e.g. sutralang.dev) are deprecated.

• Projects — https://emmaleonhart.com/projects/
• Research — https://emmaleonhart.com/research/
• Interactive ML tutorials — https://emmaleonhart.com/tutorials/

• Sutra — a geometrically compiled, GPU-native, differentiable language: vector symbolic architecture (VSA) operations as the program, compiled to tensor arithmetic, with early experiments running on thermodynamic (thrml) sampling hardware. https://sutra.yantraOS.org · source
• Yantra — an interpretable neural computer: a GPU-native operating system written in Sutra. https://yantraOS.org · source
• Loka — a Rust RDF-star triplestore with native HNSW vector indexing (vectors are just triples). https://loka.emmaleonhart.com · source
• QueryKey — a local-first social network / personal relationship manager run from your own desktop. https://querykey.emmaleonhart.com · source
• Alignment — research on steering emergently misaligned LLMs back toward alignment. https://alignment.emmaleonhart.com · source
• Latent Space Cartography — relational geometry of frozen embedding spaces. https://latent-space.emmaleonhart.com · source
• Reservoir Agent: reservoir-computing research toward a stateful, autonomous executive function for Yantra. source

• cleanvibe — Python scaffolding for agentic coding projects: the docs / queue / devlog conventions an AI coding agent works inside to stay on-task and self-documenting across long sessions. source
• Vibecoding tutorial — a beginner-friendly guide to driving AI coding agents (AI-pair-coding). repo

Know well:

Working knowledge: