A 5-Layer AI Governance Framework Built on Cursor

Introducing the story behind the development of Aria ( https://aezisai.co.jp/products/aria ).

Transcript

1. A 5Layer AI Governance Framework Built on Cursor Riu Tokiwa Aezisai Japan March 2026 PROJECT ARIA aezisai.co.jp Copyright (c) 2026 Aezis AI Inc. All rights reserved.
2. PROFILE Who I Am Riu Tokiwa / Aezisai My 2026 Workload Delivered simultaneously by one person: 4 Own Products 3 Contracted Projects 10 Parallel Workstreams Managed & Developed by 1 Person Powered by Cursor IDE Phase 2: Engineer & Educator (2016〜2019) Web developer, programming school instructor MVP award) Founded 80-person freelance guild, trained beginners to full-stack in 6 months Phase 3: PM & DX Strategist (2019〜2022) PM for listed companies, PM conference speaker Founded Aezisai 2020, DX consulting for multiple enterprises Phase 4: AI Solution Architect → "One-Person SI" (2022〜) 5 own products + 3 enterprise projects = 10 parallel workstreams, 1 person Powered by Cursor + ARIA Copyright (c) 2026 Aezis AI Inc. All rights reserved. Phase 1: Entrepreneur & Creator (〜2016) Beauty career: 15 years, 30-staff salon management Independent: video production, TV CM, 1,000-person event staging, DTP/logo design
3. CURSOR USAGE STATS Cursor 2025 Wrapped Joined 507 Days Ago Tokens 5.11B 5,106,352,451 processed Global Rank Top 3% Elite Power User Streak 128 Days Continuous Coding Agents 20.2K Sessions Initiated Tabs 277 Avg Active Context 2025 Wrapped Summary Copyright (c) 2026 Aezisai. All rights reserved.
4. CURRENT PERFORMANCE 2026 Current Performance Jan - Mar 2026 AI CODE SHARE 76.5% Record High Automation Primary Model: Claude 4.6 Opus (via Cursor) MONTHLY EDITS 813K Consistent high-volume output maintained through ARIA automation assistance. ARIA Active 5Layer Governance on Every Session Status ENFORCED Coverage 100% Sessions Latency <20ms overhead PARALLEL WORKSTREAMS 10 Simultaneous Managed by 1 Person Copyright (c) 2026 Aezisai. All rights reserved.
5. THE PROBLEM/THE SOLUTION Air-Gap Challenge & ARIA Japanese Enterprise Context The Air-Gap Barrier Mandatory Isolation Japanese enterprises require strictly air-gapped operations. No external cloud connections allowed. Governance Policy Even registering a Cursor account violates corporate governance policies for 3rd party AI tools. "The tool I loved most couldn't enter the environments that needed it most." ARIA  Cursor 5Layer Framework ARIA runs inside Cursor, solving governance gaps that were outside Cursor's current focus. Enterprise Approved Deliverables built on our devices using the ARIA  Cursor stack have been approved by clients. "We crossed the air-gap barrier without giving up Cursor." Copyright (c) 2026 Aezisai. All rights reserved.
6. Intelligent Context Assembly Instead of dumping the entire repository (~2,800 files) into the LLM's context window—which dilutes attention and spikes costs—ARIA treats the codebase as a knowledge graph. It selectively loads only the nodes strictly relevant to the active cursor position. Solves "Lost in the Middle": Prevents critical governance rules from being ignored due to context saturation. Synaptic Path Resolution ARIA uses a Rust-based graph traversal to find the shortest dependency path between the user's current task and required governance rules. This mimics human intuition: recalling specific knowledge only when triggered. Zero-Latency Governance Shifts context curation from LLM to a local deterministic runtime. This enables high-frequency usage of Opus-class models without hitting rate limits. 90% Token Reduction <50ms Retrieval Time 100% Local Execution CURSOR CONTEXT READY Local-first calculation ensures 0ms cloud latency ARIA RUNTIME REPO (~2,800 FILES) TRIGGER RULE MATCH SEC POLICY OPTIMIZED CONTEXT CORE CONCEPT Synaptic Inheritance Dynamic Context Assembly 06
7. ~2,800 Total Governance Files 9 Git Submodules Independently Versioned Design Philosophy Layered architecture ensures separation of concerns. Cursor handles the core editing, while ARIA layers handle enterprise governance, knowledge, and safety. .aria/ Layer 5 Local Project Context & Runtime Config Local Context .well-known/ Layer 4 Knowledge Base (Tech Stack, Domain Logic) 2,200+ Files 9 Categories .synaptic/ Layer 3 Governance, Meta-Cognition & Experience Rules 105 Experience Rules .ide/ Layer 2 Fundamental Architecture Principles 41 Principles .cursor/rules/ Layer 1 Native Integration Layer (The Foundation) 13 Always-Active 86 Triggers SYSTEM ARCHITECTURE 5-Layer Architecture ARIA + Cursor Stack
8. "By the time Cursor receives context, it's already curated, optimized, and safe." Cursor Context Ready Curated & Optimized Rust CLI aria check aria sync aria bdd Build-Time Enforcement Validates 5-layer rules before AI context assembly. Rejects non-compliant code patterns instantly with zero latency overhead. CI/CD Integration Runs natively in pipelines to block dangerous commits. Zero-overhead validation ensures enterprise safety standards. ZERO OVERHEAD Local-First Device-Level Pre-Learning Indexing and context optimization happen entirely on local hardware. No sensitive code leaves the device during prep. Experience Sync Updates local knowledge graphs from team repositories without cloud LLM dependency. Maintains air-gap integrity. Security Eliminates 3rd-party API risks for governance logic. Secure by design. NO CLOUD DEPENDENCY MCP Server Native Integration Plugs directly into Cursor via Model Context Protocol. Exposes ARIA's 2,800-file governance layer as a standard tool context. Protocol Benefits Standardized context delivery enables seamless "toss-up" to Claude 4.6 Opus. Context is formatted perfectly for LLM consumption. SEAMLESS CONTEXT RUNTIME & EXECUTION The Runtime Layer Local-First Architecture Copyright (c) 2026 Aezis AI Inc. All rights reserved.
9. SOLVED ARIA CORE Hallucination Elimination Zero Tolerance rules enforce strict accuracy. No made-up APIs or libraries allowed. Dangerous Command Guardrails Proactive blocking of risky operations like global git resets or rm -rf. Enterprise Automation Complex governance workflows executed reliably via ARIA protocols. THE FRONTIER FUTURE Graphical Understanding The gap: Frontend UI, draw.io diagrams, Unity structures, HTML layouts. UI/UX Unity Diagrams "I still use Genspark for UI reviews — I want to do everything in Cursor. If Cursor could understand visuals as higher-resolution source code, use cases would explode." The Opportunity Visuals as "High-Res Source Code" is the key to unlocking the next 10x utility. STATUS REPORT What's Solved & The Frontier Achievements vs. Aspirations Copyright (c) 2026 Aezis AI Inc. All rights reserved.