Internal Knowledge Workspace

In Progress

An AI-powered knowledge system that answers complex questions about our deals and operations by querying a structured data warehouse and an indexed document corpus.

Started
September 2025
Stack
Python, FastAPI, BigQuery, Firestore, Gemini, Google Cloud Run, Vercel, Next.js

The Problem

Critical institutional knowledge lived in siloed places: inboxes, spreadsheets, individual people's heads. Getting an answer to a question required knowing who to ask. The goal was a system where anyone could query the knowledge without needing historical context or SQL skills.

The connective layer across our systems. It translates natural-language questions into structured queries, pulls from the right data sources, and returns context-rich answers.

The foundation of this is boring and unglamorous: the data has to be clean before any AI layer can be useful. That’s why the email routing agent came first. A misfiled document is invisible to a knowledge system, and invisible means useless.

The system pulls from three places:

  • A structured deal graph in BigQuery — deal data, demographics, supply, competitive context, all normalized and linked
  • An indexed document corpus in Firestore with vector embeddings — every appraisal, feasibility study, and underwriting memo, chunked and citable
  • Live operational data from the underwriting platform and other internal tools

The design principle I keep coming back to: AI is the interface, not the source of truth. Every answer the system returns is traceable to a specific row, document page, or calculation. When AI is grounded in verifiable data, its output becomes indisputable.

Essays about this build