The systems behind the demos.
Now templates you own.

The flagship offer: all four Python modules — the Eval Harness, the Guardrailed MCP Server, the Safe Agent Scaffold, and the Cohere RAG Pipeline — for one flat price. Bought à-la-carte that's $446; the bundle is $299, a 33% saving. They're built to compose: the RAG pipeline exports eval cases straight into the judge, and the MCP server + agent scaffold share the same guardrail posture.

A rubric-driven LLM-as-judge with multi-axis / gated scoring, per-run cost, latency (p50/p95) and verdict metrics persisted to SQLite, prompt-drift detection, and Anthropic + generic-HTTP + deterministic-passthrough adapters. Prove your AI's answers are good — measured, not assumed.

A production MCP server: Streamable HTTP + SSE, bearer / API-key auth with scoped read/write, a typed @tool registry (JSON Schema derived from Pydantic), and a fail-closed guardrail pipeline — rate-limit → allowlist → validate → execute → audit — with a structured audit log. Plus a Copier scaffold to stamp out new servers.

A safety-first agent runtime: ReAct + Plan-and-Execute behind one interface, a confidence-floor gate, async resumable SQLite-backed human-in-the-loop approval, prompt-injection defenses (untrusted-data delimiting + scanner, AST-only safe_eval), and Sentry span-per-step tracing. Ships a documented threat model and a SOC-triage case study inline.

A RAG template with an Embedder / Reranker / Synthesizer provider abstraction — real Cohere embed-english-v3.0 + rerank-english-v3.0, a deterministic offline stub provider, a pgvector or in-memory store, configurable chunking, and a recall@k retrieval-eval that exports cases in a judge-harness-compatible format (it pairs naturally with M1).

An additive auth layer: the web app authenticates via a Clerk session cookie; an external MCP / API client presents a long-lived Clerk JWT-template bearer token; both resolve to the same Clerk user id. Middleware short-circuits bearer requests before the cookie gate. ~110 lines, heavily commented — the best-documented small asset in the set.

callClaudeJSON<T>() — a memoized Anthropic client that calls Claude, concatenates text blocks, strips ```json fences and prose, and JSON.parses the result, with a typed AnthropicError for deterministic fallback. Zero domain coupling; a single-file library.

The two halves of a RevenueCat subscription integration, extracted from a shipped iOS/Android app. Client: a Capacitor/React module with an idempotent configure guard, offerings fetch with backoff, purchase / restore / entitlement helpers, and a store-review-safe example paywall. Server: a FastAPI webhook that maps RevenueCat's lifecycle events onto durable entitlement state — constant-time shared-secret auth, a swappable entitlement store, and application that is idempotent and safe against out-of-order delivery. The webhook is the half most integrations skip, and it's the half that decides whether a paying user keeps access.

The full Safe Agent Scaffold runtime — ReAct + Plan-and-Execute behind one interface, a confidence-floor gate, resumable human-in-the-loop approval, prompt-injection defense, a resumable SQLite store, and span-per-step tracing — vendored in-tree, plus a READ-scoped retrieval tool surface that speaks the Model Context Protocol to any conforming RAG MCP server. It's a superset of the Safe Agent Scaffold, not a lighter alternative: the entire runtime ships intact. Retrieved chunks are treated as untrusted input, the tool surface is pinned to two READ-only tools, and it carries no write key. Runs fully offline on a stub brain + a recorded MCP transport — no network, no keys.