Revamp Engine

The challenge.

Real estate builders often have slow, outdated websites that fail to convert high-intent buyers. However, rebuilding custom landing pages manually takes days. Furthermore, naive AI automation pipelines that call LLMs synchronously crash completely when API providers throttle or experience transient downtime.

The solution.

Engineered an asynchronous queue-first worker architecture separating AI generation from ingestion, rendering, and deployment. Built an autonomous agentic worker pipeline that discovers builder sites, extracts property listings, generates optimized copy, compiles modern Astro templates, and holds previews for mandatory human verification.

Under the hood.

Asynchronous event pipeline where Redis holds prioritized worker queues (q:scrape, q:import, q:ai, q:deploy) backed durably by SQLite with Write-Ahead Logging. Circuit breakers isolate LLM rate limits while allowing scraping and static page builds to continue uninterrupted.

The implementation.

A durable, queue-driven pipeline where SQLite is the source of truth and Redis is a rebuildable hot cache (queues, state, circuit breaker), rehydrated from SQLite on boot. Work is segregated across dedicated per-queue workers: q:scrape, q:import, and q:deploy drain regardless of LLM health, while q:ai stalls only when the LLM circuit breaker is OPEN, ensuring a third-party model outage degrades one capability instead of the whole system. A FastAPI control plane dashboard runs continuously on port 8080 and never returns 500 status codes even if Redis is unreachable. A Node build service (buildsvc) renders production Astro templates (astro-basic and neo-brutalism) and deploys preview builds to Vercel or Netlify. A human stays in the loop by design: every revamp stops at the built state and waits for explicit human review and approval before anything deploys outward. Runs 24/7 on a Raspberry Pi 4B node.

Engineering highlights.

• Engineered queue-driven architecture where SQLite persists state and Redis serves as an ephemeral hot cache rehydrated on boot.
• Implemented automated circuit breaker pattern around LLM API calls to isolate model outages from ingestion and deployment queues.
• FastAPI control plane dashboard designed with graceful degradation to never return 500 status codes even if cache layer becomes unreachable.
• Human-in-the-loop validation stage preventing automated external publishing without manual review and sign-off.
• Low-power self-hosted deployment running continuously on a single Raspberry Pi 4B node.