Partnership Thesis
Built on Claude.
Aligned with Anthropic.
Sovereign Matrix is a multi-tenant AI workforce platform with 137 specialised agents. Claude is the premium-quality lane in our unified 8-provider router — and the model whose safety research shaped how we built the platform itself.
01 — Where Claude lives in our stack
Three Claude tiers. One unified router.
We route Claude when quality demonstrably matters. The router absorbs ~80% of inference on free providers (Cerebras, NIM, Ollama); Claude stays reserved for the workloads where verification, code correctness, and reasoning depth justify the cost.
War Room verifier
Claude Opus 4.7
Three independent agents debate a high-stakes business decision. Opus moderates the debate, scores each position, and synthesises the strongest answer. The accuracy of the verifier matters more than its cost.
Code agents
Claude Sonnet 4.6
code-agent, code-reviewer, code-sandbox all route to Sonnet. The trade-off — Sonnet beats every cheaper model on code reasoning at a price point that's acceptable for SMB SaaS pricing.
Routing + classification
Claude Haiku 4.5
Fast classification decisions inside the platform itself: which agent should handle this prompt? what category does this lead fit? Haiku at $1/M input is the cheapest reliable classifier we've benchmarked.
02 — The safety pipeline
Five layers, inspired by Constitutional AI.
Every agent output runs through a 5-layer pipeline before reaching the caller. The layer order and the critic-review pattern were shaped by Anthropic's safety research direction.
- 01
Jailbreak detection
Block prompt injection BEFORE the handler runs. Pre-flight check on input, not post-flight.
- 02
Content safety
Topic + harm filters on inbound user content. Same rationale as a moderation API.
- 03
Handler execution
Agent runs. Output goes to the next layer, never directly to the user.
- 04
PII redaction
Post-flight scan for emails, phone numbers, credit cards, SSNs in agent output. 5 patterns; small overhead.
- 05
Quality scoring + critic review
Output below threshold gets regenerated once. Critic agent reviews if useCritic=true. This is the layer Constitutional AI taught us how to think about — model-self-correction without infinite loops.
Implementation lives in src/lib/agent-factory.ts. All five layers run in parallel via Promise.all so total user-perceived latency stays under 200ms.
03 — The open-source layer
Router extracted as MIT-licensed package.
@sovereign/ai-router is the multi-provider router that powers Sovereign Matrix. Zero deps, MIT license, drops into any TypeScript project. Routes across Anthropic, OpenAI, Cerebras, NVIDIA NIM, Groq, DeepSeek, and local Ollama with one call.
import { createRouter } from "@sovereign/ai-router";
const router = createRouter({
providers: [
{ name: "anthropic", apiKey: process.env.ANTHROPIC_API_KEY },
{ name: "cerebras", apiKey: process.env.CEREBRAS_API_KEY },
{ name: "nvidia-nim", apiKey: process.env.NVIDIA_NIM_API_KEY },
],
budget: { dailyCapCents: 500, onCapReached: "fallback-to-free" },
});
await router.complete({
prompt: "...",
priority: "best", // routes Claude first
userId: "user_abc", // for budget tracking
});04 — What we'd love from Anthropic
Three concrete asks.
Anthropic Startup Program
We've applied at anthropic.com/startups. The credits would directly fund the Claude Sonnet code-agent + Opus verifier roles where we route paid traffic.
Computer Use API early access
We have a computer-use agent in the registry — currently a stub. Pilot access would let us ship a first-class autonomous-browser agent for Sovereign customers.
Co-marketing case study
Once the first SA mid-market customer ships measurable outcomes on Sovereign + Claude, that's a case study Anthropic could co-publish (emerging-markets distribution, agency replacement, vertical specialisation).
For Anthropic recruiters reading this:
Sovereign Matrix is built solo by Christiaan de Wet in collaboration with Claude as pair-programmer, per Anthropic's candidate AI guidance. Repo is public. README documents specific architectural decisions. Glad to defend any of them live, unassisted.