AI Voice SaaS Platform

Multi-tenant AI receptionist SaaS — answers business calls in under a second and completes real tasks mid-call: books appointments, quotes prices, texts confirmations, transfers to humans. 24/7.

What I Did

• →Built a multi-tenant AI receptionist SaaS that answers business calls and completes real tasks mid-call — booking appointments, quoting real prices, texting confirmations, and transferring to humans when needed.
• →Engineered the voice pipeline on Twilio + Vapi (Deepgram STT) with tool-call webhooks served by a FastAPI voice engine inside Vapi's 5-second budget — every fact (dates, prices, availability) must come from a tool call, structurally preventing hallucination.
• →Designed the low-latency data path: PostgreSQL (AWS RDS) as system of record with org-scoped multi-tenancy and raw-SQL migrations, Redis caching call state for sub-millisecond mid-call reads, plus distributed rate limiting and cluster-wide circuit breakers.
• →Replaced the vector database with hierarchical LLM tree retrieval — TF-IDF narrowing + gpt-4o-mini reasoning over section summaries (~500ms) — cheaper and more accurate for tenant-sized knowledge bases.
• →Shipped AWS-native background processing (SQS FIFO post-call pipeline with idempotent critical steps, EventBridge cron → Lambda sweeps) on ECS Fargate via GitHub Actions, and built the Next.js 15 customer dashboard and SEO marketing site.