Building “Delphine”

The Challenge

A pre-revenue startup was tackling an ambitious problem: building an AI-powered risk intelligence platform for critical infrastructure. The founding team of 10 professionals was distributed across time zones, each contributing 5–7 hours weekly during the due diligence phase.

The coordination challenge was significant. With team members working asynchronously on market research, competitive analysis, ICP validation, and financial modeling, knowledge was fragmenting across individual work sessions. Meeting notes lived in scattered documents. Insights from one workstream weren’t informing others. The project manager was spending disproportionate time synthesizing updates and ensuring alignment.

They needed a way to create shared institutional memory without adding administrative overhead.

The Approach

Rather than implementing a complex enterprise knowledge management system, I designed a custom GPT called “Delphine” to serve as the team’s strategic advisor and knowledge hub.

The setup was deliberately simple. Using ChatGPT’s shared Projects feature, we created a collaborative workspace where all 10 team members could access the same AI assistant. We established “keystone documents” as sources of truth: team charters, decision logs, research memos, and weekly updates from each workstream.

The key insight was structural, not technical. Each team lead maintains a running document with their meeting notes and weekly status updates. These connect directly to Delphine. As they update their documents, the AI’s knowledge base updates automatically. This created a living system rather than a static repository.

Delphine’s instructions emphasized two functions: surfacing existing team work product when relevant, and providing strategic analysis that connected dots across workstreams.

The Results

Within weeks, the team reported meaningful shifts in how they worked:

One team member captured it simply: “It’s getting really fun to work with Delphine.”

Why This Worked

Three factors made this implementation successful:

• Right-sized technology. A shared ChatGPT Project with custom instructions was sufficient. No enterprise software, no complex integrations, no IT involvement. The team was operational within days.
• Structure before AI. We invested time upfront defining which documents would serve as sources of truth and how team leads would maintain them. The AI amplified good information architecture rather than compensating for chaos.
• Team visibility. The shared project meant everyone could see each other’s conversations. This created unexpected value: team members learned from each other’s prompting approaches and built confidence in AI collaboration by watching peers succeed.

The Broader Lesson

This project demonstrated something I see consistently: the highest-value AI implementations often aren’t the most technically sophisticated. They’re the ones that solve real coordination problems with appropriate tools.

For a distributed team with limited hours and no shared infrastructure, a custom GPT became the connective tissue that made async collaboration feel coherent. The AI didn’t replace human judgment. It made human judgment more informed by ensuring relevant context was always accessible.

The team is still early in their venture, but they’ve built a foundation where knowledge compounds rather than fragments. That’s the kind of AI implementation that creates lasting value.