AI Engineering · New Zealand

A LangGraph Platform for AI-Powered Location Intelligence

We built five production AI workflows on a unified LangGraph + Gemini stack, improving speed with early-return results and URL-level caching. The system supports resumable conversations, enforces governance within AI nodes, uses optimized multi-layer storage, and ensures reliability through schema-validated outputs—all accelerated by reusing Reach’s existing AI framework.

Discuss Your AI Project

Client Overview

About Our Client

A New Zealand-based market intelligence platform serving real estate businesses, retail chains, investors, franchise owners, and marketing consultants. Users range from open (anonymous) to fully authenticated accounts; the platform supports both first-time exploration and saved-analysis workflows.

Industry: Marketing

Tech Stack

LangGraph
Pinecone
LangSmith
FastAPI + Uvicorn
Python 3.12 & more.

Results

Five distinct AI agentic flows in production, all on a unified LangGraph + Gemini stack.
Perceived latency reduced via early-return pattern: users see scrape results before deep enrichment finishes.
URL-level caching prevents redundant scrapes for known businesses across multiple analyses.
WebSocket conversations resumable across reconnects and server restarts via Postgres-backed checkpoints.
NZ-only governance enforced inside the AI nodes, not as fragile post-processing filters.
Three-tier storage (PostgreSQL / PostGIS / Pinecone) — each access pattern uses the store best suited.
Schema-validated outputs at the model layer eliminate free-form JSON parsing failures in production.

Core Solution

Five LangGraph AI agentic flows, each isolated
We built five separate AI workflows on LangGraph: a core business intelligence pipeline that researches and enriches data with early-return results and background processing, an onboarding chat that combines guided flows with structured LLM outputs, a post-analysis chat grounded in stored results and streamed in real time, a Gemini-powered target market mapper with caching, and a reporting flow that generates summaries with reliable fallbacks.
Nine backend domains in strict 4-layer DDD architecture
(Routes / Controllers / Services / Repository): auth, user, analysis, business_intelligence, welcome_conversation, catchment_analysis, target_market, report, analysis_conversation. Controllers know nothing about SQL; services know nothing about HTTP. Each layer independently testable
Three-tier storage optimized per access pattern:
PostgreSQL 17 for transactional and LangGraph checkpointer state. PostGIS for geospatial, Pinecone for vector retrieval (business narratives, audience segments, geo summaries — chunked and embedded per analysis namespace).
Reliability patterns
structured-output validation everywhere; NZ-clamp safety filter embedded at the node level; LLM fallbacks on missing API key or invocation failure; non-fatal partial-pipeline persistence; background-task pattern for long-running enrichment that outlives the originating HTTP request.