Search
Posts
IntentCAD Viewer — Closing the DWG FastView Gap
Web workers, animated zoom, progress feedback, and color-coded selections. The changes that took IntentCAD's browser viewer from prototype to production-grade — benchmarked against DWG FastView.
ReadDeep Dive Part 4: Z3 Formal Verification, the Three-Layer Stack, and Claude Code as Architect
How Z3 SMT solver replaces LLM-as-a-judge with mathematical proofs, the three-layer IRSB+Moat+Scout stack, and the story of Claude Code as the lead architect across 285 commits. Grounded in the smart-contract formal-verification literature.
ReadDeep Dive Part 3: A 12-Package Nested Monorepo That Watches AI Agents for You
Inside the IRSB Watchtower: a 12-package pnpm monorepo with evidence verification, behaviour-signal derivation, risk scoring, and auto-dispute — ~500 tests of deterministic monitoring, grounded in the payment-channel watchtower literature.
ReadDeep Dive Part 2: Cryptographic Receipts and the Evidence Pipeline That Proves What AI Agents Actually Did
How the IRSB Solver creates SHA-256 evidence bundles, signs them with Cloud KMS, and posts cryptographic receipts on-chain — creating an unforgeable audit trail for AI-agent work, grounded in the audit-log integrity literature.
ReadDeep Dive Part 1: Five On-Chain Enforcers That Make AI Agent Wallets Structurally Safe
How EIP-7702 delegation, five caveat enforcers, and bond staking create defense-in-depth guardrails for AI agent wallets — with 11 contracts live on Sepolia. Read against the smart-contract security literature and the agent-tool-use research.
ReadDeep Dive Part 4: Building 10 Production Gems with Claude Code as Tech Lead
What it means when the AI makes the architectural decisions. How Claude Code served as tech lead across 10 Ruby gems with approximately 2,924 tests and 60+ canonical docs — read against the LLM-agent and software-engineering-automation literature.
ReadDeep Dive Part 3: The Observability Loop — Teaching AI Tools to Improve Themselves
How the wild ecosystem's three-repo pipeline — telemetry, transcript normalization, and gap mining — creates a feedback loop that teaches AI tools what they are struggling with. Read against Dapper, fire-and-forget supervision, and the privacy-engineering literature.
ReadDeep Dive Part 2: CLAUDE.md — The Missing Manual for Human–AI Software Collaboration
How per-repo CLAUDE.md files act as binding contracts between human architects and AI implementers. Read against the literature on long-context attention degradation, sycophancy, and module boundary information distribution.
ReadDeep Dive Part 1: The Safety Architecture of Letting AI Agents Touch Your Production Rails Database
How the wild ecosystem makes it structurally impossible for AI agents to cause damage when accessing production Rails databases. Defense in depth, adversarial testing, and hard safety ceilings — read against the security-engineering and access-control literature.
ReadWrite Once, Publish Everywhere: Building Content Distribution Across Three Sites
Three sites now share content through automated pipelines: Hugo to Astro backfill, cross-repo plugin sync with repository_dispatch, and a CodeRabbit-caught no-op filter.
Read