Proof of what's possible.

Internal Tooling

Building a Private Knowledge Graph That Actually Answers Questions

The Challenge

Knowledge workers build up a mountain of documents over time. Pitch decks, reports, emails, meeting notes, spreadsheets, screenshots. After a couple of years, there are thousands of files. Folders stop working. Search misses anything that was worded differently than you remember. The information you need exists somewhere in the pile, but you cannot find it in the thirty seconds before a call.

Off-the-shelf tools do not fix this. Notion and Claude Projects do not automatically pull in new files from your inbox or Drive. Enterprise tools like Glean cost thousands per seat and require a vendor to customize. ChatGPT with file uploads forgets everything the moment you close the chat.

What was needed was a private memory system that runs on your own machine, ingests your whole document pile automatically, and lets you ask questions that pull from across all of it.

What We Built

A local system that watches a folder, pulls in every new document automatically, and organizes the contents into a searchable memory. You ask a question in plain language and get an answer that draws from across your entire document pile, with links back to every source it used.

Every document is scanned for the important stuff: the people, companies, tools, and projects it mentions. Those get organized into a graph so you can see at a glance that three different documents all reference the same person, or that five projects all depend on the same vendor.

When you ask a question, the system searches two ways at once. It searches for documents that match the meaning of your question, and it searches for documents that contain the exact words. Then it combines both results and gives you an answer with links back to the source documents so you can verify what it said.

Everything runs locally. No data leaves the machine except for the AI calls that read and summarize text, and those never touch the raw files during search. The system handles plain-language questions about risks, decisions, people, or any topic that appears across the document pile, and answers with citations attached to every claim.

The same approach works for any business drowning in documents: staffing agencies, insurance brokerages, and marketing agencies all deal with the same problem at different scales.

Results

Functional Prototype

Centralized Intelligence Platform for Venture Capital

The Challenge

Early-stage venture capital funds operate across a patchwork of disconnected tools. Deal sourcing lives in spreadsheets and bookmarked portfolio pages. Company research happens across dozens of browser tabs. Investment theses are written from scratch in Word documents. Founder outreach is managed through individual email threads.

For a lean fund team, this fragmentation creates compounding problems. Sourcing coverage is limited by human bandwidth. The research-to-decision cycle is slow, with 4 to 8 hours of manual web research per company. And institutional knowledge is lost between team members because it lives in individual files rather than a shared, searchable system.

What We Built

A locally-hosted multi-agent orchestration platform that automates the end-to-end venture capital deal sourcing workflow. The system uses specialized AI agents, each scoped to a specific domain with its own tools, memory, and reasoning capabilities, coordinated by a master agent through a web-based dashboard.

The platform has five integrated components. A Portfolio Intelligence Scanner that monitors 138 VC firms and has discovered over 800 portfolio companies through automated extraction. An Automated Research Engine that dispatches 27 specialized tools to evaluate a company in 5 to 7 minutes, including parallel data gathering, sector-adapted traction research, rigorous market sizing, competitive landscape mapping, and automated fact-checking. An Investment Thesis Generator that synthesizes research into professional 11-section documents matching institutional quality standards. An Outreach Management System with human-in-the-loop approval that ensures no email sends without explicit sign-off. And a Pipeline Dashboard that serves as the operational center for the entire workflow.

Each component feeds directly into the next: portfolio scans populate discovered companies, evaluations produce theses, theses inform outreach decisions, and all activity is tracked in a persistent, auditable record. The platform uses intelligent model routing across tiers for cost control; our guide on how to choose the right AI model for your business covers the decision framework, and our transparent breakdown of AI costs explains the economics.

Results