A lot of pages in this category prove themselves with a wall of tool logos. These are systems I built and operate myself, and two of them you can read line by line.
GulLabs is the open-source side of Gulify. These are published, licensed, and yours to inspect before you decide whether to trust me with your systems.
Apache-2.0 · npm @gullabs/any-llm
A TypeScript client for model calls, built as nine packages with a hard rule at every boundary. No credential is ever read from the environment by accident. An unknown model is rejected rather than quietly remapped to a default. Every call returns an exact cost that is never retroactively repriced. A broken logger can never take down the call it was logging.
This is the repository to open if you want to know how I will treat your systems.
MIT · a plugin for Claude Code and Codex
A production-readiness audit for software written by AI agents, built as twelve skills across UX, frontend, backend, database, security, infrastructure, observability, testing and AI integration. It runs entirely inside your own session: no network calls during an audit, no telemetry, nothing transmitted anywhere. The companion that applies fixes is deliberately forbidden from marking its own fixes as resolved, because a tool should not grade its own homework.
The most recent release shipped because it found a live unauthenticated payment webhook in production that TypeScript, ESLint, thousands of passing tests and a running error monitor had all missed.
A workspace for agencies running social media at scale: approval workflows, multi-platform publishing, client reporting. Multi-tenant by design, with every query scoped to a workspace, and long-running AI work on durable workflows rather than hopeful HTTP requests. The prompt engineering lives in a separate, access-restricted repository, so developers building the product never see the AI logic they are consuming.
The publishing engine behind OneCast: ten platforms, one interface, and three consumption surfaces from a single codebase, a REST API, a scheduling portal and an internal library. Around three thousand tests on the core library alone. Failures are classified into four kinds, so a dead token, a malformed post, a transient error and a rate limit are each routed to the right remedy automatically instead of all surfacing as failed.
A luxury talent and photography platform, built as three applications and three published packages. The AI engine that produces its editorial content lives in a separate repository with its own database and talks to Veloir only through a versioned SDK, so either side can change without breaking the other. That discipline is why your automation will not be welded to whichever model is fashionable this quarter.
An AI picture-book studio for ages two to seven. Every stage of the generation pipeline is a pure, replayable function behind a documented quality gate, so a bad result can be traced to the stage that produced it and re-run, rather than regenerated and hoped over.