The problem
Matera started as property-management software and now runs property management itself. Its property managers answer their clients’ questions all day, and every answer sits somewhere in 6M+ documents.
Finding it by hand took about thirty minutes a question. That time, more than anything else, capped how many clients each manager could take on.
What I built
I was the lead GenAI engineer and owned the whole chain behind the assistant: roadmap & strategy, alignment with stakeholders, cloud infrastructure, ingestion, search, evaluation and production.
An ingestion pipeline sized to the budget. Fully processing the whole corpus would have cost about six times more. Instead, a lighter pipeline makes every document reachable, and the assistant searches them agentically: it finds and reads what each question needs. That costs a sixth as much, without losing accuracy.
Knowledge search with an assistant on top. Managers ask in plain language; the assistant searches the documents and emails and answers with what it found.
Agents connected to the application’s data, through MCP. With the tech team, I gave the agents access to the application’s own data, read-only for a start, so an answer can combine a document with the live record it refers to. I built part of the MCP tools myself, search among them, and integrated the MCP server the tech team runs in the Rails app for wider data access.
Evaluations of the agents’ tool calls. I evaluated how the agents call the application’s data with a 200-question Langfuse test dataset, plus manual analysis of their traces in Langfuse.
AI spend under control. I routinely ran FinOps on the Vertex AI projects I owned and cut their spend by half.
Owned by the team. Matera’s product runs on Ruby on Rails; the GenAI layer is Python, beside it. I worked with the managers on what they needed and with the engineering team on how it fit, and the product’s Rails developers contributed to the Python codebase.
Results
- Managers find an answer in 5 minutes instead of 30, and save 1 to 2 hours a day each: about 100 hours a day across the team.
- Each property manager can now take on more clients, so Matera can scale its managed property service. Its VP Data & AI calls the project “crucial for the company’s future”.
- ~100 people use it every day, and usage grew 60-fold in nine months.
