Just Now Possible with Teresa Torres
Episode: Building Earmark: How a Two-Person Team Turned Meetings into Finished Work
Date: February 5, 2026
Host: Teresa Torres
Guests: Mark Barbier (Co-founder & CEO, Earmark) and Sanon (Co-founder & CTO, Earmark)
Episode Overview
This episode explores the creation and evolution of Earmark, a productivity AI tool developed by a two-person founding team to transform meetings into actionable deliverables for product builders. Teresa Torres hosts a deep-dive conversation with Mark Barbier (CEO) and Sanon (CTO), covering the impetus behind Earmark, the specific pain points it aims to solve, its unique technical architecture, real-world usage, and the lessons learned as they iterated on both product and workflow.
Key Themes and Insights
1. Defining the Problem: Meetings That Create More Work
-
The "Infinite Workday"
Mark references a Microsoft Research paper defining the "infinite workday": back-to-back meetings, endless context-switching, and little time for focused work. Earmark addresses the resulting "fight or flight" response, where practitioners struggle to do their actual jobs (02:17). -
Core Customer
Earmark is tailor-made for product builders—product managers, engineers, designers—especially teams overwhelmed with the administrivia of tracking and following up on meeting outcomes (01:07, 03:02).
2. What Makes Earmark Unique vs. Other Meeting Tools
-
Beyond Note-Taking: Finished Work, Not Just Summaries
Earmark distinguishes itself by tracking not just notes but generating finished artifacts (product specs, Jira/Linear tickets, live prototypes, slides) in real time, reducing manual follow-up (03:51)."At the end of the meeting...it really attempts to give you finished work before your meeting ends." — Sanon (03:51)
-
Real-Time, Live Functionality
Multiple AI agents operate in parallel, providing live artifact generation and support, including on-the-fly translations of engineering jargon, custom compliments, and more (04:30–05:39)."As you are speaking, you actually have essentially multiple agents that are running in parallel." — Sanon (04:30)
-
Templates and Personas
A suite of live templates includes acronym explanations, "make me look smart" prompts, actionable minutes, and agent personas (e.g., Security Architect, Accessibility, Legal) that ask contextually relevant questions as if key stakeholders were present (06:01–08:12).
3. Product Evolution: From Apple Vision Pro to AI Meeting Agent
-
Early Pivot
The initial idea was an immersive AR/VR tool for communication skill-building on Apple Vision Pro—a pivot was necessary due to limited market size and actual user behaviors (17:21–19:02). -
Validation Through Usage, Not Data Storage
The first web versions didn’t store data; everything was ephemeral. This “no storage” approach surprised them by helping win over enterprise customers concerned about privacy and served as early validation. Enterprises liked evaluation-friendly, non-persistent data (16:10–16:41)."You're telling me you're not going to store or train [on] data. We couldn't even do it if we wanted to. That actually has helped us get in the door..." — Sanon (16:10)
-
Iterative Prototyping Based on Customer Conversation
Five major product iterations were made with continuous customer feedback loops, closely monitoring synchronous Slack channels and proactively adopting customer language (19:02).
4. Technical Architecture and AI Strategies
Live Meeting Processing with Multiple Agents
- Speech-to-text via Assembly AI streams every ~30 seconds.
- Delta-based transcript batching minimizes LLM costs.
- Uses OpenAI (primarily GPT-4.1 for “prose quality,” but tests various models) (23:13–25:07).
Prompt Caching for Cost Efficiency
"Prior to prompt caching, the economics of this tool were actually completely untenable—at one point it was around $70 for an hour meeting... now, it's sub a dollar." — Sanon (23:13–24:58)
- Prompt caching is essential for live, affordable operations. OpenAI now offers prompt caching similar to Anthropic APIs (28:26).
- Only the transcript is passed as conversation history; cards/templates are top-level, fresh prompts—preventing prompt pollution and optimizing cache utilization (32:11–34:30).
No Speaker Diarization
- Chose not to do live speaker attribution due to technical reliability issues—wrong names are more harmful than none (21:57).
Data and Privacy
- Ephemeral by design; data is not written to a persistent store by default, favoring enterprise privacy. Can be set globally for organizations (52:30).
5. Information Retrieval & RAG Challenges
-
Beyond Simple Vector Search
Earmark is pushing past basic vector search or keyword RAG for multi-meeting analytics and synthesis."A lot of our users are actually wanting to do more analysis-based questions...the answer doesn’t live in the transcript—RAG with cosine similarity won’t help." — Sanon (35:47)
-
Hierarchy and Pyramids of Data Borrowing ideas like the "data pyramid" (raw transcripts at the base, increasingly abstracted insights above)—to narrow search spaces and precompose likely-needed summaries (42:30).
"Ideally we can find the answers at the top of the pyramid, but if we need to, we can walk all the way down to the transcript." — Sanon (44:14)
-
Agent-Native Workflows
Implements Dan Schipper’s "agent-native architecture": not just RAG but full agents that tool around databases, keywords, metadata, and pre-computed summaries for nuanced queries (35:50–39:56).
6. Artifacts, Templates, and Customization
-
From Talking to Working Artifacts Any meeting conversation can become:
- Product specs (pushed to Jira/Linear)
- Prototypes via V0 or Cursor
- Presentation decks (“generate presentation from these three conversations, sprinkle in some emojis”)
- Detailed requirements, custom prompt-based outputs
"...what used to be a week in duration or two weeks...now it's just immediate...all the artifacts required from the kickoff ready by the end." — Mark (48:24)
-
Iterative & Inclusive Use
- Templates offer a “magic moment” MVP; power users graduate to custom prompts that better fit exact use cases (46:22–47:36).
7. Measuring and Maintaining Quality
-
Hallucination Control Explicit escape hatches in prompts (“if you don’t know, say so”) are crucial for LLMs to avoid inventing facts (49:32).
-
Trust Through Provenance Teresa’s suggestion of requiring LLMs to provide timestamped or line-referenced proof of any fact—improves both hallucination prevention and user trust (54:28).
-
Human-Centric Evals Evaluations are currently conducted by hand—looking for usage signals like artifact copying and manual review in the absence of stored data. Tech debt includes formalizing evals after establishing user habits (49:32–51:47, 56:39).
Notable Quotes & Moments
- Mission-Driven Product Building
"We can't imagine not building for people like us." — Mark (00:37)
- On Data Ephemerality’s Surprising Value
"Enterprise prospects saw that as almost a feature...‘Oh, like you're telling me you're not going to store, not going to train…’" — Sanon (16:10)
- On Their Philosophy of Innovation
“The best pivots take you back home.” — quoting Dalton Caldwell (19:02)
- Prompt Engineering Wisdom
"It's a lot better to give the model the specific content rather than just stuff it with as much context...which yields worse results.” — Sanon (35:08)
- AI as Chief of Staff Vision
“Nothing falls through the cracks. Deliverable quality is there...to provide comfort and folks feel truly supported in these servant leadership roles.” — Mark (58:29)
- UI/UX Learning
"We noticed...our users fixated on the transcript. If it misspelled a name, they'd want to edit. What if we solve this through UX: minimize it to subtitles?" — Sanon (20:33)
Important Timestamps
- [01:16] Mark describes Earmark and problem space (“productivity suite where work completes itself”)
- [03:51] Sanon explains real-time, finished work creation—difference from Granola
- [05:39] Teresa probes live agent functionality
- [13:58] Vision for multiplayer/team features
- [17:21] The original Apple Vision Pro idea and user research/pivot
- [23:13] Deep technical details—speech-to-text, transcripts, prompt caching
- [32:11] Challenges in agent prompt caching, preventing context pollution
- [35:47] Why basic RAG (vector/keyword search) isn’t enough
- [42:30] Teresa’s “data pyramid” analogy for scalable synthesis
- [44:33] Concrete example: meetings to engineering specs/prototypes
- [49:32] Evaluation, hallucinations, and privacy
- [58:29] Vision: the future AI chief of staff for product teams
Concluding Summary
Earmark stands as an example of high-leverage, customer-obsessed AI product development: A tiny team is shaping the future of work for product builders by moving from capturing meeting notes to creating real, actionable, finished deliverables in real time. Their journey—from early AR/VR experiments to a privacy-forward, AI-driven productivity suite—shows both the promise and challenge of building thoughtful human-centered tools in the age of agents, LLMs, and workflow automation.
For those building with AI:
This conversation is packed with concrete technical strategies (prompt caching, agent architectures, RAG/retrieval synthesis), product philosophy (build for “people like us”), and practical lessons (embrace ephemerality, privacy as an architecture, experiment relentlessly), making it an essential listen/read for anyone designing the next wave of AI-augmented collaborative software.
