WizAI

Routes incoming messages from WhatsApp and Instagram to a centralized AI processing pipeline, normalizing platform-specific message formats (WhatsApp Business API webhooks, Instagram Graph API events) into a unified internal message schema. Implements platform-agnostic conversation threading that maintains context across both channels for the same user, enabling seamless handoff and consistent conversation history regardless of which platform the user contacts.

Generates contextually appropriate responses using an LLM (likely GPT-3.5/4 or similar) that understands conversation history, user intent, and platform norms. Applies platform-specific formatting rules post-generation: WhatsApp responses respect message length limits and markdown-style formatting, while Instagram responses optimize for character limits and emoji usage. Implements few-shot prompting with user-provided training examples to customize response tone and domain knowledge without fine-tuning.

Automatically detects the language of incoming messages and translates them to a configured default language for AI processing. Translates AI-generated responses back to the customer's original language before sending. Supports 50+ languages using translation APIs (Google Translate, AWS Translate, or similar). Implements language-specific customization (e.g., different training examples per language) to improve response quality beyond generic translation.

Connects WizAI to external CRM systems (Salesforce, HubSpot, Pipedrive) and business tools (Shopify, WooCommerce, Stripe) to access customer data, order history, and account information. Enables AI responses to reference real-time data (e.g., 'Your order #12345 shipped on Monday') without manual data entry. Implements bidirectional sync: incoming conversations can create/update CRM records, and CRM data can be used to personalize AI responses.

Processes incoming images and videos from WhatsApp and Instagram conversations using computer vision APIs (likely AWS Rekognition, Google Vision, or similar) to extract visual content understanding. Generates contextual responses based on image analysis (e.g., 'That's a great product photo! Here's the link to buy it') or routes media to appropriate handlers (product identification, damage assessment for insurance claims). Supports media attachment in outgoing responses, enabling the AI to send images/videos back to users when relevant.

Allows users to provide conversation examples (user message + desired AI response pairs) that are stored and used as few-shot prompts in the LLM context window. Implements a simple UI or API for uploading training data without requiring technical ML knowledge. Stores training examples in a vector database or simple key-value store, retrieving relevant examples based on semantic similarity to incoming messages to inject into the LLM prompt dynamically.

Detects when an incoming message requires human intervention (e.g., complex requests, sentiment indicating frustration, or explicit 'talk to a human' keywords) and automatically routes the conversation to a human agent queue. Implements rule-based detection (keyword matching, sentiment analysis) and optional ML-based confidence scoring to determine handoff threshold. Preserves full conversation history and context when handing off, so agents see the complete interaction without re-asking questions.

Tracks and aggregates metrics on AI-generated conversations including response times, customer satisfaction (inferred from follow-up messages or explicit ratings), handoff rates, and message volume trends. Provides dashboards showing which response types are most effective, which conversations get escalated, and which training examples drive the best outcomes. Implements basic attribution to link conversation outcomes (purchase, support resolution) to specific AI responses or training examples.

Enables users to schedule messages to be sent at specific times or in batches to multiple contacts via WhatsApp and Instagram. Implements a queue-based system that respects platform rate limits and delivery guarantees, with retry logic for failed sends. Supports templating (e.g., 'Hi {customer_name}, your order {order_id} is ready') to personalize bulk messages without manual composition for each recipient.

Stores complete conversation history (all messages, metadata, timestamps) in a persistent database, enabling the AI to reference past interactions when generating responses. Implements context window management to fit relevant history into the LLM's token limit (e.g., summarizing old messages to preserve recent context). Supports conversation search and retrieval so users can find past interactions with specific customers.

Enforces WhatsApp and Instagram message template requirements (e.g., WhatsApp requires pre-approved templates for business messages, Instagram has character limits and emoji restrictions). Provides a template builder UI that guides users through platform-specific rules and automatically submits templates for approval to Meta's systems. Tracks template approval status and prevents sending messages that don't match approved templates, reducing risk of account suspension.

Analyzes incoming messages to detect customer sentiment (positive, negative, neutral) and emotional tone (frustrated, happy, confused, etc.) using NLP models or third-party sentiment APIs. Feeds sentiment scores into the AI response generation to adjust tone appropriately (e.g., more empathetic for frustrated customers, celebratory for happy ones). Flags high-priority conversations (very negative sentiment) for human review or escalation.