AI Space

Sun, 10 Aug 25

Demonstrated meeting note generation and transcript features
Screen recording through accessibility tools captures all text on screen
- Records continuously during meetings and work sessions
- Converts visual information to plain text (no OCR needed)
- Creates searchable database of all screen activity
Privacy controls available - can pause/disable recording
Beta access currently limited, requires form completion

Explained difference between basic LLM calls vs RAG systems
- Standard transcript → LLM summarization (what Granola does)
- RAG involves vector embeddings for document retrieval
- Context windows now large enough to include full transcripts without RAG
Vector embeddings process:
- Convert text to numerical vectors
- Search for closest matching vectors when querying
- Return top 5 most relevant text chunks to LLM
Memory systems create persistent user profiles from accumulated data

Frameworks vs Providers distinction:
- Frameworks: CrewAI, LangChain, LangGraph (code libraries for building agents)
- Providers: AWS Bedrock, Google Vertex AI, Azure AI Studio (API access to multiple models)
Bedrock/Vertex don’t have proprietary models - provide access to Gemini, Claude, etc.
- Single API key gives access to multiple model providers
- Leverage cloud infrastructure (AWS servers, GCP servers)
MCP (Model Control Protocol) simplifies tool integration
- Eliminates need for complex prompt engineering
- Provides structured documentation for agent capabilities