A modular AI engine with installable skills that execute real outcomes. I am building it as founder and CTO, launching Q3 2026.
Relaystack is an AI platform that executes real-world outcomes through composable, installable skills — not just conversation.
The architecture treats AI models as a runtime, not a product. Skills are discrete units of capability that combine to do things: schedule meetings, process invoices, monitor infrastructure, and report back. Each skill is versioned, testable, and replaceable.
As sole engineer, I designed the monorepo structure, the skill contract API, the queue system, and the orchestration layer from scratch — then built against them simultaneously.
Most AI products stop at the conversation. Getting an LLM to generate text is table stakes. The hard part is reliability: running skills that touch real systems, recover from failure, and audit what they did. That is the problem Relaystack is built to solve.
Skills are npm packages that implement a typed contract. They can be installed, updated, or swapped without touching the engine. The platform grows without the core changing.
Every skill execution is a BullMQ job with retry, backoff, and dead-letter handling. If the process crashes mid-skill, the job resumes — not restarts. Transient infrastructure failures do not become user-facing errors.
Complex tasks decompose into sub-agents that run concurrently, report back, and can spawn further sub-tasks. The orchestrator never blocks on a single model call, and every step is logged for auditability.
Relaystack is the platform I wanted to exist. It will be in production Q3 2026 — built to the standard I hold everything else to.