A PropTech / real-estate startup: AI property app (wiki + MCP + agent + fetchers)

In one line. 80x moved from a CRM audit into a build-and-architecture role on a UK PropTech startup's AI property app, scaffolding its knowledge wiki, redesigning its MCP-server architecture, and going deep on the fetching and memory stacks that turn buyer preferences into leads.

Client & context

The client is a UK PropTech / real-estate startup led by an experienced property-tech founder. Its app captures buyer preferences through property analysis and converts buyer leads into seller leads over time. The relationship started as a CRM audit (Airtable → Attio, $200/hr, $1k audit) and grew onto the product itself.

The problem

The core tools and a deep agent already existed, ready to shape into a production product. The knowledge base needed structure; the MCP layer was a single server without versioning or date stamps; the deep agent ran against an outdated workflow version; and the fetching stack spanned ~15,000 estate-agent sites, where scripted fetchers break and pure-LLM fetching is too variable and costly.

What we built / scoped

• A knowledge-wiki scaffold in the Karpathy style: raw, date-stamped transcripts as the source layer, with an abstracted layer that always links back.
• A three-server MCP architecture with versioning and date stamps, internal/admin (all tools), a curated agent-facing server (treating the agent as adversarial), and an external/ChatGPT-facing server exposing the app's own agent, with the deep agent to be re-wired onto the new MCP.
• A fetching strategy: a coded fetcher scored against a baseline, falling back to Firecrawl below threshold, with Moti generating reusable per-domain fetchers plus drift detection.
• Memory-system design across three schemas, personal context, search preferences, search leads, with inferred vs. confirmed memories.
• A full project backlog spanning the MCP servers, agent and UI.

How we did it

The work also reviewed the two in-house analysis tools: the floor-plan pipeline (OCR + YOLOv8 + SAM segmentation, ~90% accurate on simple two-beds) and the Django image-condition grader. In a live session we authenticated into the MCP server via ScaleKit/Google and confirmed 18 tools; a full property analysis on a live listing returned floor-plan, condition and image results and surfaced a new lead.

Outcome

The app is ~90% of the way to V1: 18 MCP tools are live and the full property-analysis pipeline runs end to end in ~5-10 minutes at ~$2 per analysis. The engagement is ongoing, with wiki scaffolding, the backlog and repo access as the immediate next steps.

Takeaway: An 18-tool MCP stack runs a full property analysis in ~5-10 minutes at ~$2 a run, roughly 90% of the way to V1, with the architecture set up to version and scale.