English · 简体中文 · Русский · हिन्दी · العربية

Founder: Dayna Blackwell. I build the open infrastructure the agentic AI stack runs on, and ship the evidence to back it: durable agents, wire formats, code intelligence, MCP tooling, and conformance testing, with the research underneath. Independent, evidence-first.

The AI-native wire format for structured data. 100% comprehension on every frontier model. 50-92% fewer tokens than JSON. 2,500+ LLM evaluations across 11 models and 4 providers. 43B+ lossless round-trips across 5 formats. Deployed in 20 production systems including Chrome DevTools MCP. Zero training required.

Spec · Go · TypeScript · Python · Rust · Swift · Kotlin · .NET · Proxy · Tree-sitter

A full framework for building AI agents in Go: models (OpenAI, Anthropic, Gemini), tools, typed multi-step flows, memory and RAG, MCP, multi-agent coordination, and typed human-in-the-loop, all on one durable, append-only journal. The journal is the difference: side effects fire at most once (a resumed run never re-charges a card or re-sends an email), thousands of concurrent runs survive crashes and node handoffs in a single process with no cluster, every run emits a cryptographically verifiable audit trail (RFC 6962 Merkle proofs, checkable without trusting the vendor), and shared governed state is provably convergent. Built for ambient agents that run unattended and act under audit.

Getting started · Concepts · Flows · Governance · Audit · Security model

Code intelligence infrastructure for AI agents. 65 tools, 30 CI-verified languages, 24 agent workflows. Single Go binary. Uses GCF as default output format.

Conformance testing for MCP servers. 102 servers scanned, 34 bugs found, 12 upstream issues filed. Fuzz testing, schema linting, per-assertion Docker isolation.

Self-adapting code intelligence engine. The system GCF was extracted from. 28 MCP tools, graph-native analysis, session deduplication.

Parallel AI agent coordination. Disjoint file ownership, git worktree isolation, tier-gated execution, and human-reviewed plans. A Scout agent maps the codebase into a coordination plan; Wave agents implement their assigned files simultaneously.

Protocol · Claude Code · Codex · Go

Go structs to TypeScript and Zod, one source of truth. Discriminated unions, enums, maps, and validation rules compile to runtime-checked Zod schemas; opt-in json-tag inference bridges plain types. Built for Wails and web frontends. Apache-2.0.

Local implementations of Google Cloud APIs for development and CI. No GCP credentials required.

Secret Manager (50K+ downloads) · KMS · IAM · Eventarc · Auth · IAM Control Plane · Core

9 self-published papers (Zenodo DOIs), several with industry response. A research program on tokenizer-attention coupling proving that BPE merge decisions permanently constrain transformer attention capacity, plus systems work on distributed convergence and memory reclamation.

Tokenizer-Attention Coupling How BPE merge decisions permanently shape transformer internal organization. 43 tokenizers, 20 providers. Controlled experiment: identical models, different tokenizer. Merge barriers produce 3-738x better structured data comprehension, zero natural language cost. 18-phase causal ablation across 2 architectures, 2 scales, 3 domains.

Stranded Attention A previously undescribed failure mode: every attention head in a standard BPE model has structural capacity the tokenizer permanently prevents. All 384 heads at 410M and 768 at 1.3B show 4x more delimiter attention under clean boundaries. The 40pp frustration gap appears by step 5,000 and never closes.

Developmental Atlas of Attention Head Specialization First head-specialization atlas at scale: 384 heads, 7 behaviors, 131 checkpoints, 7 runs, 2 architectures. The BPE capacity tax is architecture-independent (+64.3% NeoX, +67.0% Llama). 48-56% of attention capacity in standard BPE is non-productive.

Structural Ambiguity in JSON Tokenization Cross-tokenizer analysis across 8 tokenizers, 6 providers. JSON field names fuse with the opening quote on 50-63% of tokenizers. JSON boundary merge rate 8.93% vs 1.00% for pipe; TOON tab 59.82%. JSON overhead reaches 81% at 500 rows.

Graph Compact Format (GCF) The AI-native wire format for structured data. 2,500+ evaluations across 11 models and 4 providers. 43B+ lossless round-trips. Deployed in 20 production systems. Spec v3.5.1 Stable. Explore the project

The Hierarchical Identity Architecture Content-addressing as a computation primitive for software relationship intelligence. 2.75x more precise than GitNexus (p=0.0003), 193x faster indexing on enterprise repos. Explore the project · merkle-strata

Memory Drainability Formalizes when coarse-grained allocators can reclaim memory. Proves a sharp O(1) vs Ω(t) dichotomy for bounded retention. Validated empirically (238x recycle-rate differential). Explore the project

Normalization Confluence Coordination-free convergence via well-founded compensation. Third convergence regime alongside CRDTs and invariant confluence. Explore the project

Federated Normalization Confluence Multi-organizational convergence through morphism validity preservation over acyclic networks.

40+ merged PRs across the ecosystem. #6 contributor to mcp-go (8.7K stars). Data corruption fixes, panic recovery, SDK hardening, spec compliance, transport bugs.

Full list

Languages:

Platforms & Shells:

Developer Tooling:

Artificial Intelligence: