Terravive
Multi-tenant AI site-intelligence SaaS for architects, built around a multimodal RAG pipeline and LangGraph agents on a type-safe Turborepo monorepo (NestJS 11 API + Next.js 16 RSC admin).
// problem
Architects needed one multi-tenant platform to ingest and query everything known about a site (PDFs, CAD, GIS, LiDAR, photos, audio), with trustworthy AI answers, strict tenant isolation, billing, and a fully type-safe contract between the API and the admin app.
// solution
Architected a multimodal RAG pipeline: BullMQ workers (OCR, Whisper, Claude vision) ingest site data into pgvector, retrieved by hybrid vector + full-text search with RRF fusion and Cohere reranking. On top sits a LangGraph agent graph (planner, tool registry, input/output guardrails) over a multi-provider LLM layer (Claude, OpenAI, Groq) with circuit-breaker fallback, semantic caching, and per-tenant budget guards. The platform is a type-safe Turborepo monorepo where one Drizzle/PostgreSQL schema drives the OpenAPI contract and typed TanStack Query hooks, secured with JWT + OAuth, RBAC, tenant isolation, audit logging, Stripe billing, and PostHog feature flags.
// outcomes
- [✓]Retrieval quality gated in CI with golden-set evals: recall@3 1.0, MRR 0.90 on real site data
- [✓]Nightly LLM-as-judge scoring and OpenTelemetry tracing across the agent graph
- [✓]Multi-provider LLM layer with circuit-breaker fallback, semantic caching, and per-tenant budgets
- [✓]End-to-end type safety from database schema to frontend hooks, with multi-tenant isolation
next project
Stedi