A multi-portal real estate listing automation platform. It ingests brochures, images, and video, structures them with AI into typed property records, and auto-publishes listings across multiple portals from a single trigger. Built solo, end to end.
The agency listed every property by hand, retyping the same details into portal after portal, each with its own multi-step wizard, login, and quirks. Slow, error-prone, and impossible to scale as inventory grew.
One system that reads the raw media once, lets AI structure it into clean property records, then drives each portal's full publishing flow automatically. Manual distribution became a single JSON event.
Brochures, images, and videos land in Nextcloud and trigger the pipeline automatically, no manual upload step.
Each media type is parsed by the right model: OCR for PDFs, vision for images, transcription for video, then embedded for semantic search.
An LLM turns messy extracted content into clean, typed property fields, stored once in a proyecto table so extraction never re-runs per portal.
An Inngest-based execution layer coordinates the whole flow with idempotent retries, so a failure resumes cleanly instead of duplicating listings.
Stagehand + Browserbase drive each portal's full wizard: 15 stages on Corotos, 18 steps on Mercado Libre, and a custom Facebook Marketplace flow, with captcha-tolerant logins and persistent server-side sessions.
Every publish is logged with an audit trail, and Telegram alerts report success or failure in real time.
One event, full distribution. A single JSON trigger takes a property from raw media all the way to live listings across every portal, hands-off.
Residential DR proxies, persistent server-side browser context, and captcha-tolerant logins keep automated sessions alive on sites built to block them.
Durable execution with retry logic means a dropped step resumes without ever creating a duplicate listing, safe to re-run.
Structured data is written once into a shared table instead of re-running costly LLM extraction for every portal.
Each portal's multi-stage publishing flow is modeled separately, 15-stage, 18-step, and custom, with image curation per platform.
PDFs, images, and video are each routed to the right model and unified into one clean, queryable property record.
Architecture, backend, AI pipeline, browser automation, and ops, designed and shipped end to end by one engineer.
I turn messy business problems into reliable AI systems, ingestion, agents, RAG, and automation, designed and shipped solo.