~/architecture
Six architecture diagrams: a personal multi-agent OS run by Claude Code, distributed microservices, multi-tenant SaaS, and an ML product. Every node was extracted from the real codebases, not sketched.
built on Claude Code
The brain of the whole system is Claude Code. The orchestrator is an interactive Claude session that fans work out to custom Skills via the Agent tool; email watchers spawn headless claude -p workers; the mobile bridge runs the Claude Agent SDK; MCP servers wire it into the vaults. I build with Claude, not just on top of a model API.
These six diagrams were generated and verified by Claude Code multi-agent workflows: one sub-agent authored and rendered each diagram, then independent agents audited it against the real codebase and knowledge base for accuracy and reviewed the render for legibility. This page was built by the system it documents.
Every domain as a column, every shared platform drawn once as a capability band. Scan down a column to see what a project depends on; scan across a band to see who shares a platform. This is how five very different systems compose into one engine, now coordinated through a shared orchestration layer of specialist agents, saved workflows, and policy-as-code.
An always-on multi-agent operating system run by Claude Code. The orchestrator dispatches through a versioned orchestration layer (seven specialist subagents, a saved multi-agent workflow library, policy-as-code hooks) out to custom Skills via the Agent tool; systemd-timed watchers spawn headless claude -p workers; a layered memory (session → project → vault) compounds research and ops, and a weekly self-verification loop audits the docs against the live system. Local models on an RTX 3080 handle the free tier; the iOS bridge runs the Claude Agent SDK.
A polyglot, event-driven recording platform built for graduate distributed-systems work (CAS 735). Six services across Node and Python, each a hexagonal ports-and-adapters bounded context owning its own data, choreographed over a RabbitMQ topic exchange plus a dedicated work queue for FFmpeg media processing.
Four roles (member, instructor, admin, platform owner), each with distinct authentication and capabilities, gated through a single pipeline: role-scoped JWT → edge middleware → multi-tenant subdomain routing → app-level studio isolation. The isolation boundary is enforced in the data layer by a scoped-Prisma wrapper, not by database RLS.
The same product traced by data movement: a synchronous request spine (booking, auth, card checkout) split cleanly from a durable, event-driven async tier (payment webhooks with HMAC verification, Inngest functions, scheduled cron, realtime push). One shared serverless Postgres, every row studio-scoped.
A resume-tailoring ML product with paying users. A JWT-gated request enqueues a durable, step-checkpointed Inngest pipeline that drives the OpenAI Responses API with Zod-validated structured outputs. Shown sparse on purpose: only the publicly self-reported stack, nothing inferred.
M.Eng research project
Mac Study Companion was my M.Eng Software Engineering research project at McMaster, supervised by Dr. William Farmer and Dr. Richard Paige, built for neurodivergent learners. It turns lecture recordings into study notes with a guarantee attached: every sentence it releases carries a citation back to the minute of audio it came from, and any sentence the checker cannot prove is held back in place with its reason shown, never quietly dropped. A fail-closed release gate discharges the per-response verification conditions in Z3, and the pipeline state machine is model-checked in TLA+.
Every design decision traces to a 25-item evidence-graded ADHD learning-design checklist distilled from the literature, in the spirit of Nielsen’s heuristics, and strict enough that it caught the product’s own quiz flow revealing correctness on the same click as the pick, which was rebuilt into answer-then-reveal. Both supervisors accepted the project in September 2026 as sufficient for the degree, and the build concluded in October 2026.