A real-estate technology company wanted every one of its agents to have their own AI assistant. That meant running a fleet of roughly 70 customer machines, each with a live AI agent, plus the AI features to make them useful, and a way to keep the whole thing observable and cheap to run.
An AI assistant per agent sounds simple until you have to run 70 of them for real customers. Three problems stacked up: the assistants needed genuine AI features (market analysis, lead handling, not a toy chatbot), the fleet had no visibility into what each agent was costing or whether answers were any good, and standing up a new machine was a slow, manual, error-prone chore. The brief was to make the whole platform reliable, observable, and fast to grow.
I built it as a system with clear layers: data comes in, agents reason and act through tools, the fleet runs the agents, and everything reports back to one observability layer. Provisioning feeds new machines into the fleet on demand.
Intent recognition with LangExtract, action-plan documents generated as a microservice, and integrations into Gmail, Telegram, and GoHighLevel. Grew from v1.6 to v2.0 in production.
LLMs plus computer vision over live MLS data, an image-classifier microservice on Cloud Run, email-reply feedback loops, and automated workflow tests.
A plugin on each box streams token, cost, and quality data to a zero-dependency relay, then to a Supabase registry that self-enrolls machines. API keys stay off customer machines. Cost-by-model and error rates land in one dashboard.
Ansible plus a baked DigitalOcean image took a new machine from about 34 minutes to about 16, validated end to end in production.
Migrated LLM models across 41 live machines with zero failures using idle-gating and automatic rollback, and hardened data security with Supabase row-level security and Cloud-Function token encryption.
The platform went from an idea into a running product: agents that do real real-estate work, a fleet you can see and trust, and a growth path that does not fall over when the next customer signs up.
Book a free call. You will leave with a clear plan for your own AI automation, whether or not we work together.
Book a free call