I build tools that make AI agents waste fewer tokens and remember more.

Helioy is an ecosystem for AI-augmented software engineering. The thesis is simple: context windows are the bottleneck. Every token should be high-value. Navigation is waste. So I'm building the infrastructure to eliminate it.

attention-matters — Geometric memory engine on the S³ hypersphere. Persistent recall for AI agents using quaternion drift, phasor interference, and Kuramoto coupling. One brain per developer, not per project. Rust.

fmm — Structural metadata for source code. Auto-generated sidecars that tell LLMs what a file does without reading it. 80-90% fewer file reads. Rust + tree-sitter.

mdcontext — Structural intelligence for markdown. Hybrid search (BM25 + semantic), section-level indexing, 80%+ token compression. TypeScript + Effect.

nancyr — Multi-agent orchestrator wrapping Claude Code CLI. Hub-and-spoke coordination, token budgets, adaptive TUI. Rust.

helioy-plugins — Claude Code plugin bundling skills, MCP servers, and hooks. The glue that connects everything.

nancy — The shell prototype that proved the architecture. Process supervision, token management, hook servers. Where it all started.

Three libraries that each solve one problem well:

attention-matters  →  memory      (what happened before?)
fmm                →  code        (what does this file do?)
mdcontext          →  documents   (what does this doc say?)

One orchestrator that composes them:

nancyr  →  spawn agents, compile context, coordinate work, learn from outcomes

All exposed as MCP servers. All usable independently. All designed to make the next AI interaction cheaper and more accurate than the last.

• Ship working code. Fix forward.
• CLI wrapping over direct API. Meet developers where they are.
• One brain, not one brain per project. Cross-pollination is a feature.
• No magic numbers. Constants derive from phi and pi. The math decides what matters.
• Independent repos, not a monorepo. More scalable, independent brands.