A personal AI assistant platform structured as a super-app — one identity, four focused services, every AI call routed through a single gateway I built to handle provider routing, cost metering, and audit logging.
I wanted a single place for the AI tooling my family and I actually use — household finance, Indonesian stock-ownership intelligence, vacation planning — without lock-in to one model vendor, without losing track of what each feature is costing me, and without trusting opaque AI calls in production code I run myself.
Built it as a super-app — like Gojek or Grab, one identity gives access to several focused services — with every AI call across every app flowing through a central gateway (CMS Maya). The gateway handles provider routing (Anthropic ↔ Gemini ↔ OpenAI per-feature via SQL, no redeploys), cost metering in USD per call, versioned prompts, and a tamper-proof audit hash chain. Two-database split keeps consumer data isolated from ops + audit data. Deterministic logic does the heavy lifting; AI only makes judgment calls, defaulting to the cheapest viable model per feature.
Four apps live and in daily use — Maya Buddy (household finance with AI receipt scanning), Maya Finance (BEI 1%+ shareholder tracker with AI PDF auto-parse + Q&A), Maya Elysia (cuti optimizer + AI trip concierge), and CMS Maya (the gateway + admin dashboard). Projected ~$30/mo at scale. 192 automated tests across the suite surfaced 3 production bugs pre-ship — including a rate-limiter that failed closed.