Undertow

Undertow is a self-reported decision-audit tool for teams that processes meeting transcripts or recordings to extract discrete decisions with owners, categories, assumptions, and confidence levels. It compares new decisions against prior ones in a workspace and flags contradictions, supersessions, reaffirmations, or refinements. The product is described as an MVP built using FastAPI, React, Clerk, GPT-5.6, GPT-OSS-20B, Whisper, and other technologies.

The description states that Undertow aims to be more useful than a meeting summarizer by creating a "decision ledger" rather than a chat summary. It includes features like workspace scoping, notifications, and tension review. The authors claim they built a complete path from source meeting to reviewable audit record, including synthetic test fixtures.

The single most important open question is: What is the actual commercial traction or adoption of Undertow beyond this hackathon MVP?

This analysis is based entirely on the self-reported project description provided by the authors. No independent verification, revenue data, customer names, or usage metrics are available.

The description states that Undertow:

• Accepts uploaded transcripts or browser recordings
• Transcribes recordings using Whisper
• Extracts discrete decisions with owner, assumption, category, and confidence
• Compares new decisions against prior ones in a workspace
• Flags relationships between decisions (contradiction, supersession, reaffirmation, refinement)
• Provides a "decision ledger" not a chat summary
• Includes features like ledger view, meeting archive, recording flow, tension review, notifications, categories, profile, workspace views
• Uses SQLite for local storage and Clerk for authentication

The authors describe the system as having a single pipeline for both transcripts and recordings after transcription.

Inference: The product appears to be a decision-tracking tool that uses LLMs to extract structured data from meeting content and compare it over time. It is not described as a calendar or scheduling tool, nor does it appear to integrate with existing platforms beyond authentication.

The description states:

• Teams "rarely lose the whole meeting. They lose the decision inside it."
• Undertow is positioned as being "more useful than a meeting summarizer"
• It aims to be a "decision audit" that keeps a "durable record of commitments"
• The tool is described as creating a "decision ledger, not a chat summary"

The authors claim that Undertow:

• Keeps a durable record of commitments
• Flags when that record changes
• Inspects source meetings and explains why relationships were flagged
• Allows resolution or reopening with notes
• Maintains an immutable review history

Inference: The positioning is that Undertow is a tool for teams to maintain accountability over decisions made in meetings, rather than just capturing what was said. It's positioned as a solution to the problem of "decision drift" or "unresolved commitments."

The description states:

• The product targets teams who have meetings where decisions are made
• It addresses the issue of "nobody can say what was agreed, why, or whether a later meeting quietly changed it"
• The tool is designed for teams that want to audit their own decision-making process

Inference: The target customer appears to be collaborative teams (likely in tech or business settings) who need to track decisions over time and maintain accountability. It's not described as targeting individual users, but rather groups or organizations.

The description does not contain any information about:

• Revenue model
• Pricing structure
• Monetization strategy
• Customer acquisition costs
• Unit economics

Not evidenced

The description states:

• Built with FastAPI backend (SQLite, Pydantic validation)
• Frontend built in React, TypeScript, Vite
• Uses Clerk for authentication
• Groq for live MVP
• GPT-OSS-20B for structured decision extraction and tension reasoning
• Whisper-large-v3-turbo for transcription
• Codex with GPT-5.6 used during development
• Local recording storage
• Full prior decision history passed into reasoning calls
• No embeddings or retrieval infrastructure at this scale

Inference: The product is described as a small-scale MVP built using modern web and LLM technologies, with a focus on structured data extraction and comparison logic.

The description states:

• The product is an MVP
• It includes synthetic test fixtures (Meridian fixture with contradiction/supersession/reaffirmation; Undertow Labs demo fixture with 15 meetings, 30 decisions)
• The authors claim to have built a "complete path from source meeting to reviewable audit record"
• No mention of real users, customers, or adoption beyond the hackathon

Not evidenced

The description does not contain any information about:

• Competitors
• Market size
• Competitive advantages
• Differentiation strategy
• Industry landscape

Not evidenced

The description states:

• The hard part was deciding what counts as a decision, not extracting text
• They had to stop the ledger from becoming a list of every sentence said
• They used few-shot examples, strict JSON schemas, Pydantic validation, and synthetic ground-truth sets to keep the pipeline honest
• A key challenge was teaching the system the difference between contradiction and intentional change of direction

Inference: Key risks include:

1. Difficulty in accurately identifying decisions from unstructured speech/text
2. Risk of false positives or negatives in decision extraction
3. Limited scalability without retrieval infrastructure or embeddings
4. The product is described as a hackathon MVP with no commercial traction

1. What specific use cases drove the development of Undertow?
2. How do you plan to validate that your decision-extraction logic works reliably across different meeting types and domains?
3. Are there any real-world teams currently using Undertow or testing it in production?
4. What are the key assumptions about user behavior regarding decision auditing?
5. How will you scale beyond the current MVP architecture (e.g., retrieval, background jobs)?
6. What is your go-to-market strategy for reaching target customers?
7. Have you considered how to handle edge cases like ambiguous or incomplete decisions?