Feta

Automatically captures audio streams from Zoom, Microsoft Teams, and Google Meet via native platform integrations or browser-based recording, then applies speech-to-text processing (likely using cloud-based ASR engines like Google Speech-to-Text or Whisper) to generate full meeting transcripts. The system handles variable audio quality and multi-speaker scenarios by normalizing input before transcription, enabling downstream processing of meeting content without manual recording setup.

Processes full meeting transcripts through a large language model (likely GPT-4 or similar) with a specialized prompt engineering pipeline that extracts summaries, key decisions, and action items in a single inference pass. The system likely uses few-shot prompting or fine-tuning to understand meeting context (project names, participant roles, business domain) and avoid generic verbose summaries, producing structured outputs that distinguish between decisions, action items, and discussion points.

Provides APIs and webhook endpoints to export meeting summaries, transcripts, and action items to external tools (Slack, email, project management platforms) via standardized formats (JSON, CSV, or platform-specific APIs). The system likely implements a webhook-based push model for real-time distribution and a pull API for on-demand retrieval, with support for custom field mapping to adapt Feta's output schema to downstream tool requirements.

Automatically identifies and labels speakers in meeting transcripts using a combination of audio fingerprinting (voice biometrics) and meeting metadata (participant list from platform APIs). The system likely maintains a speaker profile database keyed by voice characteristics and meeting context, enabling consistent speaker attribution across multiple meetings and reducing manual speaker labeling overhead. Role inference (e.g., 'client', 'team member', 'manager') may be derived from meeting metadata or historical patterns.

Indexes all meeting transcripts and summaries using vector embeddings (likely OpenAI embeddings or similar) to enable semantic search across the meeting library. Users can query with natural language (e.g., 'What did we decide about pricing?') and the system returns relevant meeting segments ranked by semantic similarity, rather than keyword matching. The system likely maintains a vector database (Pinecone, Weaviate, or similar) indexed by meeting date, participant, and topic for efficient retrieval.

Aggregates meeting data (duration, participant count, talk time distribution, action item completion rate) into a dashboard that provides team-level and individual-level insights. The system likely computes metrics asynchronously (daily or weekly aggregation jobs) and caches results in a time-series database for fast dashboard rendering. Insights may include trends (e.g., 'meeting duration increasing over time') and anomalies (e.g., 'participant X rarely speaks in meetings').

Implements a freemium model where users can capture and summarize a limited number of meetings per month (likely 5-10) without payment, with automatic tier upgrades triggered by usage thresholds. The system tracks usage metrics (meetings captured, API calls, storage) and presents upgrade prompts when users approach limits, enabling low-friction onboarding and conversion to paid tiers. Pricing tiers likely correspond to meeting volume (e.g., 'Starter: 10 meetings/month', 'Pro: 50 meetings/month').