Brayan
AI Agent Engineer
The problem
Most AI personal assistants depend on a paid API and live in their own app. Brayan started from a different need: an agent you talk to in natural language, directly in WhatsApp, that handles real tasks — sending an email, logging an order, setting a reminder for a specific time — without a recurring bill to an LLM provider.
The hard part isn't wiring a model to WhatsApp. It's designing a safe harness: an agent with access to your real email and real contacts needs strict authorization, not just a well-written prompt.
The solution
Brayan splits responsibilities across two processes: a Node bridge (whatsapp-web.js) that only moves messages between WhatsApp and a webhook, and a FastAPI (Python) orchestrator that holds the agent logic — short-term memory per conversation thread, a tool registry with function-calling, and a scheduler for recurring tasks.
The brain is Hermes 3 (Nous Research) running 100% local via Ollama, with an optional fallback to Anthropic or OpenAI when more capability is needed. Development follows a spec-driven cycle: every new capability starts as a spec with acceptance criteria, gets implemented with TDD, and closes with an adversarial verification pass before being marked done.
Security isn't a layer bolted on at the end: only the authorized number (AUTHORIZED_WHATSAPP_ID) can trigger tools, and if that variable isn't set the agent responds to no one — it fails closed, not open. Self-programming (the agent scheduling its own reminders) is limited to explicitly allowed actions, with no arbitrary code execution.