Skip to content
← All workDeployed

AI Property Brochure Generator

Architected a multimodal copy-generation service that fuses listing data, vision-model photo captions and geospatial place search into style-governed brochure text, delivered as serverless Azure Functions behind API Management.

Role
Solution design · AI architecture & prompt design
Year
2024–2026
Type
Full-stack platforms
Stack
Azure OpenAI (vision + reasoning) · Azure Functions · TypeScript · Azure API Management · Google Places API · React · TypeScript · Node.js · Express · Cosmos DB · Entra ID · MSAL

The problem

A real-estate consultancy's marketing staff wrote every brochure's location and property copy by hand to a strict house style. The data already existed in the property database and photo store; the challenge was assembling it reliably and writing in a consistent, guardrailed voice.

Architecture

  1. 01

    Lookup

    A property reference resolves address, coordinates, attributes, features and listing photo URLs.

  2. 02

    Enrich

    Google Places nearby search per category keeps top-rated places, with haversine distances and a wider radius outside the capital.

  3. 03

    Caption

    Each listing photo goes to a vision-capable model at low detail for a short brochure-style caption, failing soft to null.

  4. 04

    Generate

    Layered prompts combine role, style guide, places, filtered metadata, captions and guardrails; a second pass rewrites the draft against the style guide.

  5. 05

    Serve

    An Express host validates Entra ID tokens against JWKS and proxies to Azure Functions through API Management.

Engineering highlights

  • 01

    Layered, guardrailed prompting

    Prompts compose role, a house style guide with good and bad examples, geospatial context, filtered metadata, photo captions and style-only few-shot examples, plus explicit guardrails on sensitive details, length and output language.

  • 02

    Two-stage generation

    On the GPT path, drafts are rewritten against the style guide alone; the reasoning-model path skips the rewrite and uses its own request shape. Per-stage timings are returned to the UI.

  • 03

    Serverless re-architecture

    Evolved from an Express prototype to an authenticated, rate-limited, tested API, then to TypeScript Azure Functions behind API Management with a thin Express host handling Entra ID validation and static hosting.

  • 04

    Context-preserving regeneration

    Original prompts are cached under a UUID with a 12-hour expiry, so user-steered rewrites reuse full context; vague instructions return the original text unchanged instead of drifting.

Next projectAI-Assisted PMO Platform