Besty AI

Analyzes incoming WhatsApp messages using LLM-based abstractive summarization that preserves conversation context and speaker intent. The system integrates directly with WhatsApp's message stream via webhook/API polling, processes messages asynchronously to avoid blocking chat flow, and returns summaries inline or via bot responses. Handles multi-turn conversations by maintaining a sliding window of recent messages to preserve narrative coherence across long threads.

Detects and processes conversations mixing multiple languages and code-switching patterns (e.g., English-Spanish-Hindi in single message) using language identification models that tag each token/phrase with its language before passing to the LLM. The system maintains separate context for each language pair and applies language-specific prompting to preserve meaning across code-switched boundaries. Supports 50+ language combinations including low-resource languages often missed by generic LLMs.

Processes images shared in WhatsApp conversations using computer vision models (likely CLIP or similar multimodal embeddings) to extract text, objects, and semantic content. Images are uploaded to Besty servers, analyzed asynchronously, and results returned as text descriptions or structured data (OCR text, object labels, document type classification). Supports document types including receipts, invoices, screenshots, and photos with specialized extraction pipelines for each.

Automatically categorizes and tags WhatsApp conversations using LLM-based classification that understands conversation topics, urgency, and project context. The system analyzes message content, sender patterns, and conversation history to assign tags (e.g., 'urgent', 'project-x', 'vendor-negotiation') and organize chats into folders or priority levels. Tags are applied asynchronously and can be manually refined by users to improve future classification.

Collects messages from specified WhatsApp chats over configurable time windows (hourly, daily, weekly) and generates consolidated digests that summarize activity, highlight key decisions, and list action items. The system uses time-aware summarization that groups messages by topic and temporal clusters, then applies multi-document summarization to create coherent digests. Users can configure digest frequency and receive summaries via bot message or external notification.

Implements webhook-based message interception that captures incoming and outgoing WhatsApp messages in real-time, routes them to Besty's processing pipeline, and returns AI-generated responses or metadata back to the chat. The system uses WhatsApp Business API webhooks (or proprietary polling for personal accounts) to receive message events, processes them asynchronously in a queue-based architecture, and injects bot responses back into the conversation stream. Handles rate limiting, message ordering, and delivery guarantees.

Tracks user interactions with AI-generated summaries, tags, and responses to learn preferences over time. The system uses feedback signals (manual tag corrections, summary edits, response ratings) to fine-tune prompt templates and classification models through in-context learning or lightweight fine-tuning. Maintains per-user preference profiles that influence summarization style (verbose vs. concise), tag taxonomy, and response tone.

Manages LLM context limitations by maintaining a sliding window of recent messages and automatically summarizing older messages into compressed context. When conversation history exceeds the LLM's context window (typically 4K-8K tokens), the system summarizes messages outside the window into a condensed summary that preserves key facts and decisions, then includes this summary in the prompt alongside recent messages. This allows analysis of arbitrarily long conversations without losing historical context.

Implements configurable message retention and deletion policies that allow users to control how long Besty stores message data and analysis results. The system supports automatic deletion of messages after configurable periods (24 hours, 7 days, etc.), per-chat retention policies, and manual deletion requests. Provides transparency into what data is stored, how it's used for model training, and allows users to opt-out of training data usage.