Building Earmark: How a Two-Person Team Turned Meetings into Finished Work

Teresa Torres interviews the co-founders of Earmark, a productivity suite that runs multiple AI agents during live meetings to produce finished artifacts such as specs, tickets, and prototypes rather than passive summaries. The conversation covers the founders' pivot from an Apple Vision Pro presentation coach to a web-based meeting assistant, and why an ephemeral no-storage architecture became an enterprise selling point. It also details engineering tradeoffs, including cutting per-meeting AI costs from about $70 to under $1 via prompt caching, and why vector search struggles with analytical questions across many meetings. The episode is useful for PMs seeking to reduce follow-up work and for builders exploring real-time AI architectures.

01Key takeaways

• Judge AI meeting tools by whether they produce finished deliverables, not just summaries nobody reads.
• Running several specialized agents in parallel during a live conversation can surface perspectives a single summary would miss.
• An ephemeral, no-storage design can become a sales advantage when enterprise customers worry about data retention.
• Prompt caching can reduce real-time LLM costs by orders of magnitude, making high-frequency AI features economically viable.
• Vector search is insufficient for analytical questions; combining retrieval methods such as keyword, metadata, and summaries works better.
• Designing for product managers as an extreme user can yield a tool that serves a broader audience.

02Key sections

Beyond Summaries

Earmark positions itself against generic AI notetakers by generating usable deliverables during the meeting itself. The founders describe the product as a suite where the work completes itself.

Parallel Agents and Personas

Template-based agents such as an engineering translator and an acronym explainer run alongside personas that simulate absent stakeholders like security, legal, or accessibility reviewers. This lets users surface perspectives they would otherwise miss.

Pivot and Ephemeral Architecture

The team moved from an Apple Vision Pro presentation coaching tool to a web-based assistant. Choosing not to store meeting data became a trust advantage with enterprise buyers.

Making Real-Time AI Affordable

Prompt caching and model selection cut per-meeting costs dramatically, and the founders explain why an older model still produced better prose for their use case.

Agentic Search Across Meetings

Vector search alone fails on analysis questions spanning months of conversations, so the team combines multiple retrieval tools including keyword search, metadata queries, and bespoke summaries.

03From the post

“Listen to this episode on: Spotify | Apple Podcasts What if your meetings could actually produce the artifacts you need—specs, tickets, slides—before the call even ends? In this episode of Just Now Possible, Teresa Torres talks with Mark Barbir (CEO) and Sanden Gocka (Co-Founder), the co-founders of”

04Frameworks mentioned

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.