PRODUCT_VISION.md

Live Conversational Threads - Product Vision

Status: Draft Last Updated: 2025-11-10 Owner: Product Team

Mission Statement

Create a privacy-first, AI-assisted conversation analysis tool that helps humans have better conversations by:

1. Tracking parallel insights that emerge during discourse
2. Decomposing implicit claims (factual, normative, worldview)
3. Facilitating real-time with intelligent nudges and suggestions
4. Respecting privacy through open-source, self-hosted architecture

Core Philosophy

The Alignment Problem as Product:

"I want the model to chunk my linear conversations JUST the way I would segment it. That is where the alignment aspect comes in."

This project serves dual purposes:

1. Product: Improve human conversation quality through AI assistance
2. Research: Test alignment approaches on a concrete, measurable task (conversation segmentation)

Competitive Positioning

vs Google Recorder

Feature	Google Recorder	LCT
Privacy	Cloud storage, closed source	Self-hosted, open source
Platform	Pixel only	Platform agnostic
Data Ownership	Google owns	User owns
Speaker Analytics	Basic	Deep (bandwidth, interruptions, cruxes)
Conversation Facilitation	None	Real-time nudges, suggestions
Multi-source	Audio only	Calls, chats, social media
Claim Analysis	None	Factual/Normative/Worldview taxonomy

Unique Selling Propositions (USPs)

1. Privacy-First: Own your data, open source, self-hostable
2. Parallel Thread Tracking: Never lose a tangent again
3. Claim Decomposition: Understand the implicit worldviews in discourse
4. Real-time Facilitation: AI as conversation coach, not just recorder
5. Multi-source Aggregation: Unified view across all communication channels
6. Obsidian Integration: Seamless knowledge management workflow
7. Personal Alignment Training: Fine-tune on your own segmentation preferences

Feature Categories

Tier 1: Foundation (MVP)

Current state + Google Meet transcripts

• Live audio streaming with chunking
• Conversation graph visualization (structural + contextual)
• Bookmarks and contextual progress markers
• Obsidian Canvas export/import
• Google Meet transcript import with speaker diarization (ADR-001)

Tier 2: Core Differentiators (Next 3 months)

Features that set us apart

• Parallel Thread Management (ADR-002)

  • Track tangents as they emerge
  • Suggest retrieval during conversation lulls
  • Queue of pinned topics with opportunity cost visibility

• Claim Taxonomy System (ADR-003)

  • Factual claims (verifiable)
  • Normative claims (ought statements)
  • Worldview assumptions (implicit ideology)
  • Automatic classification with confidence scores

• Speaker Dynamics Dashboard (ADR-004)

  • Bandwidth hogging metrics (who talks most)
  • Interruption patterns
  • Turn-taking balance
  • Speaking time heatmaps

Tier 3: AI Facilitation (3-6 months)

Real-time conversation assistance

• Conversation Lull Detection (ADR-005)

  • Identify natural pauses
  • Suggest tangent retrieval at appropriate moments
  • Score "richness potential" of dormant threads

• Crux Detection (ADR-006)

  • Identify agreement/disagreement points
  • Surface underlying assumptions causing divergence
  • Suggest clarifying questions

• Real-time Suggestions

  • Tangent connections ("This relates to [earlier topic]")
  • Fact-checking requests
  • Steelmanning unclear statements
  • Fallacy flagging

• Agent-Assisted Research

  • Spin out background agents to fetch facts
  • "Prayer for memes" fulfillment
  • Context injection at appropriate moments

Tier 4: Meta-Conversation Features (6-12 months)

Help groups optimize their discourse norms

• Goal Tracking & Drift Detection

  • Set conversation intentions (clarify, bond, learn X, get advice)
  • Measure progress toward goal
  • Alert when rambling or quality-per-token drops

• Local Norm Optimization

  • Detect interruption culture
  • Suggest norm adjustments
  • A/B test facilitation strategies

• Shared Transparency View

  • All participants see the same graph in real-time
  • Collaborative thread management
  • Democratic "pin topic" voting

Tier 5: Ecosystem Integration (12+ months)

Network effects and multi-platform

• Multi-source Aggregation

  • Import from Slack, Discord, Twitter DMs, email
  • Unified conversation history across channels
  • Cross-platform thread linking

• Obsidian Deep Integration

  • Auto-create notes from transcription
  • Fuzzy search across all conversations
  • Bidirectional linking with existing vault
  • CRM functionality

• Legal/Professional Domains

  • Compliance-ready audit trails
  • Deposition analysis
  • Meeting minutes automation
  • Action item extraction

Known Issues & Risks

Silent Failures

1. Audio-to-text accuracy: No user visibility when transcription is wrong
2. Hallucinated summaries: LLM may invent topics not actually discussed
3. Bandwidth stats missing: Current version lacks speaker participation metrics

Technical Debt

1. Live streaming and transcript processing use different code paths
2. No A/B testing framework for segmentation quality
3. Limited observability into LLM decision-making
4. No user feedback loop for improving segmentation

Privacy & Ethics

1. Recording consent flow unclear
2. Speaker identification may reveal PII
3. Claim classification could introduce bias
4. Facilitation suggestions might manipulate discourse

Success Metrics

Product Metrics

1. Adoption: 1000 weekly active users within 6 months
2. Retention: 60% 30-day retention
3. Engagement: Average 3 conversations analyzed per user per week
4. Quality: 80% user satisfaction with segmentation accuracy

Alignment Research Metrics

1. Segmentation Accuracy: >90% agreement between user and LLM chunking
2. Fine-tuning Effectiveness: Personal model outperforms base by 15%+
3. Claim Classification: >85% accuracy on factual/normative/worldview taxonomy
4. Interpretability: Users can explain why LLM chose each segment boundary

Target Users

Primary Personas

1. The Researcher (You)

• Alignment researcher exploring concrete alignment problems
• Needs fine-grained control and interpretability
• Willing to tolerate rough edges for cutting-edge features
• Values privacy and data ownership

2. The Facilitator

• Professional meeting facilitators, coaches, mediators
• Needs real-time assistance during conversations
• Values speaker dynamics insights
• Wants to improve group discourse quality

3. The Knowledge Worker

• Researchers, writers, consultants
• Has many conversations across multiple platforms
• Needs to extract insights and action items
• Wants Obsidian integration

4. The Privacy Advocate

• Uncomfortable with Google/corporate data collection
• Willing to self-host
• Values open source
• Technically capable

Secondary Personas

5. The Legal Professional

• Lawyers, paralegals, compliance officers
• Needs audit trails and searchable records
• Values accuracy and reliability
• Has budget for professional features

6. The Team Lead

• Manages distributed teams
• Needs meeting analytics and action tracking
• Wants to improve team communication patterns
• Values transparency

Development Principles

1. Measurable Alignment

Every feature should have a clear alignment analog:

• Segmentation accuracy = Value alignment accuracy
• Claim decomposition = Understanding human values
• Facilitation = Helpful, harmless, honest assistance

2. User Feedback Loops

Build in continuous user correction:

• Allow manual re-segmentation with explanations
• Collect user preferences on facilitation suggestions
• A/B test prompts and show users the differences

3. Privacy by Design

• Local-first architecture where possible
• Encryption at rest and in transit
• Clear data retention policies
• Easy data export and deletion

4. Open Source Everything

• Core functionality open source
• Transparent algorithms
• Community contributions welcome
• Optional paid hosting/enterprise features

5. Iterative Deployment

• Ship MVPs quickly
• Gather user feedback
• Iterate based on real usage
• Don't wait for perfection

Roadmap Timeline

Q1 2025 (Current)

• ✅ Live audio streaming
• ✅ Graph visualization
• ✅ Obsidian Canvas integration
• 🚧 Google Meet transcript support (ADR-001)

Q2 2025

• Parallel thread management (ADR-002)
• Claim taxonomy system (ADR-003)
• Speaker dynamics dashboard (ADR-004)
• Privacy architecture refinement

Q3 2025

• Real-time conversation facilitation
• Crux detection
• Lull detection and retrieval
• Personal fine-tuning experiments

Q4 2025

• Goal tracking and drift detection
• Multi-source aggregation (Slack, Discord)
• Obsidian deep integration
• Public beta launch

2026

• Legal domain features
• Network effects and sharing
• Mobile apps
• Enterprise features

Open Questions

1. Fine-tuning approach: RLHF vs supervised learning vs preference learning for personal alignment?
2. Real-time latency: How fast must facilitation suggestions be to be useful?
3. Facilitation ethics: When does helpful suggestion become manipulation?
4. Business model: Open source + paid hosting? Freemium? Enterprise licensing?
5. Scaling architecture: When to migrate from GCS to distributed system?
6. Community vs product: Should this be a tool or a platform?

References

• ADR-001: Google Meet Transcript Support
• The Alignment Problem
• Discourse Quality Metrics
• Privacy by Design Principles