AI-powered personal finance SaaS — tech co-founder in a team of four, from zero to shutdown. Users send "spent $50 at the grocery store" on WhatsApp and the system categorizes, records, and analyzes in real time. Co-founded with a team of complementary profiles — as the tech person, I owned the entire stack: from architecture to deployment, across payments, AI, and conversational channels.
Two major versions — what I learned between them was the most valuable part.
Each message type was a PHP class with the action name, semantic description, and structured prompt: field schema, Portuguese examples, and rules for handling ambiguity. OpenAI returned structured JSON for the backend to validate and execute.
Async jobs with run polling, tool call execution, multi-provider orchestration (OpenAI and DeepSeek), per-message cost tracking, and abstraction via contracts (AIOrchestrator). Event-driven and Assistants API were new territory — I learned both through spec-driven development with AI, writing specs before writing code.