2026 – present
Portfolio MCP
The shortest path between a question and its answer.
- Role
- Designer and lead engineer
- Hardest problem
- Most managed MCP servers are overly complex and expensive. Our team does highly specialised analysis that requires specific optimisations.
- Result
- AI models can answer complex analytical and research questions using our data and domain knowledge.
The question
Portfolio managers have questions all day, every day. AI models are now strong enough to answer many of these questions when given access to the right data. The trouble is that MCP servers are expensive, slow, and bloated. I wanted to give portfolio managers a trusted connection to the data.
The approach
It’s an internally distributed MCPB extension, written in Go, using MCP. The data itself lives in BigQuery. Striking a balance between good context and minimal bloat is a challenge and one I continue to fight every day. Working on a small team let me act as both the subject matter expert and the developer, giving precise context while leaving room for open-ended questions.
I chose not to force or expose tool calls for specific analytics. In my analysis, these almost always required more calls and produced slower, less precise results.
All tool calls boil down to raw SQL. Once we pin down a few metrics we like, we can export those queries directly to our analysis for consistent, verified reporting workflows.
Designed to be light
The goal was low latency and low overhead:
- queries are shaped and pruned before they reach BigQuery, so each one scans only the data it needs,
- frequently used results are cached close to where they’re needed,
- the model is given only the context it needs, which keeps responses quick and costs modest.
What I took from it
Often the most useful thing I can do is remove a step. Here, the step was the wait between having a question and seeing the evidence, and removing it changed how people work with their own data.