ArgoFetch // Databode · Bodey AI assistant
Streaming AI property intelligence grounded in live Supabase queries
ArgoFetch is a production real-estate intelligence platform built for the Sunshine Coast property market under the Databode brand. The platform serves real-estate agents and investors who need fast, reliable answers about local sales, listing volumes, price trends, and time-on-market data. The product's core is Bodey — an AI assistant that answers natural-language questions about property data using a live Supabase database, not pre-computed summaries.
Bodey is sold as a SaaS product with tiered subscription plans, usage quotas, and an operator-facing admin panel. The system had to resolve natural-language timeframes precisely, enforce strict data attribution, and handle complex multi-tool queries in a single streaming response — all on top of complete SaaS infrastructure: tiered billing, per-user monthly quotas, a configurable operator admin, an interactive property map, and a module system that lets the operator change behaviour without code changes.
The product surface
Coralsoft built ArgoFetch as two linked Next.js 16 applications — the customer-facing platform and the ArgoFetch operator dashboard — sharing a single Supabase instance. The Bodey chat endpoint runs on Edge runtime for zero-buffering streaming; all other API routes run on Node.js. Every Bodey request travels through a seven-stage pipeline — auth & quota validation, chat session resolution, dynamic system-prompt assembly from Supabase, tool registry, real-time tool execution, post-stream cleanup, and usage accounting.
Bodey AI assistant
Streaming chat over OpenAI GPT-4o via the Vercel AI SDK, with five Zod-typed tools running live Supabase queries mid-response — enabling chained reasoning and multi-step analysis without leaving the stream.
Five-tool data registry
analyzeSales, analyzeListings, analyzeMatchedListings, buildChartSpec, and getDataSourceInfo — each a typed contract over the live property dataset, with a timeframe resolver mapping phrases like "last financial year" or "Q3 2023" into structured ranges.
Inline chart generation
buildChartSpec returns Recharts-compatible visualisation specs that render inline after the response completes — so chart metadata is extracted from the stream and hydrated client-side without a second roundtrip.
Interactive property map
Google Maps API integration with property pins, suburb overlays, and a configurable analytics module system — operators can drag, drop and reconfigure dashboard layouts without redeployment.
Stripe billing + quotas
Tiered subscription plans with per-user monthly prompt quotas enforced at the auth stage of every Bodey request — immediate rejection if exhausted, no wasted tokens.
Operator admin dashboard
A standalone Next.js 16 app sharing the same Supabase instance — AI instructions, pricing tiers, data imports, module layouts and attribution rules are all editable from the admin UI without code changes.
Decisions, not just dependencies.
Two Next.js 16 applications share a single Supabase Postgres instance — the customer platform and the operator dashboard — communicating through Supabase Auth with SSR cookies and Bearer-token API access. The Bodey chat endpoint runs on Edge runtime to keep streaming latency near zero; every other Route Handler runs on Node.js, where the seven-stage pipeline (auth, session, prompt assembly, tool registry, tool execution, post-stream cleanup, accounting) lives.
The AI behaviour itself is database-driven: system prompts, attribution rules, formatting logic, and module layouts are assembled from Supabase at request time, so operators can change Bodey's behaviour without a deploy. Tool calls execute live SQL against Postgres during the stream — no pre-aggregated tables, no caching layer between question and answer beyond an in-process TTL cache for hot read paths.
What was hard, and how we shipped it.
- 01
Seven-stage streaming AI pipeline
Auth & quota validation, chat session resolution, dynamic system-prompt assembly from Supabase, tool registry + model invocation, real-time tool execution against live SQL, and post-stream cleanup (chart metadata extraction, title generation, usage accounting) — all running through one streaming response on Edge runtime.
- 02
Live tool-call data retrieval, not pre-aggregation
Five Zod-typed tools execute live Supabase queries mid-stream — analyzeSales, analyzeListings, analyzeMatchedListings, buildChartSpec, getDataSourceInfo — enabling chained reasoning across the property dataset without ever falling back to a precomputed summary.
- 03
Operator-driven behaviour, no redeploys
Every operational layer — AI instructions, pricing tiers, data imports, module layouts, attribution rules — is editable through the operator admin dashboard. New behaviour ships without a deploy and without a developer in the loop.
- 04
Dual-app architecture on one Supabase
Customer platform and operator dashboard ship as two independent Next.js 16 apps but share a single Postgres database with Row-Level Security per app surface — no duplicated schema, no cross-service drift, no auth bypass.
- 05
Inline chart generation from the AI response
buildChartSpec returns Recharts-compatible visualisation specs as part of the streaming tool-call output. Charts render inline once the stream completes — no second round-trip, no separate analytics endpoint.
Looking at a project that sits at this kind of seam?
Bring us the architecture, the constraints, and the ship date. We will bring the rest.