Stammer

Provides a drag-and-drop interface for agencies to construct conversational AI flows without writing code. The builder likely uses a node-based graph system where agencies connect intent recognition, response generation, and API call nodes to define chatbot behavior. Responses are powered by underlying LLM inference (model selection unclear from available data), with visual state management replacing traditional prompt engineering and code deployment.

Enables agencies to deploy chatbots under their own brand identity without exposing Stammer infrastructure or branding. This likely involves customizable UI theming (colors, logos, fonts), domain mapping (custom subdomain or embedded widget), and client-facing analytics dashboards branded with agency colors. The deployment architecture probably uses containerized instances or multi-tenant isolation with per-client configuration overrides.

Enables chatbots to support multiple languages, with automatic language detection and response translation. The platform likely detects user language from initial message and routes to language-specific response templates or uses LLM-based translation. Agencies can define responses in multiple languages or rely on automatic translation, with language-specific knowledge bases and intent definitions.

Provides tools for agencies to review conversation logs, identify failure cases, and iteratively improve chatbot performance. The platform likely surfaces low-confidence conversations, user feedback, and intent misclassifications, allowing agencies to add training examples, refine intent definitions, or adjust response templates. Changes are deployed without downtime, and performance improvements are tracked over time.

Provides workspace and permission management for agencies to organize multiple client chatbots, assign team members to specific clients, and control access levels (admin, editor, viewer). The platform likely uses role-based access control (RBAC) with per-client isolation, allowing agencies to manage billing, usage, and team assignments at the client level. Agencies can invite team members, set permissions, and track usage per client.

Allows agencies to upload client documents (PDFs, web pages, FAQs, product documentation) which are chunked, embedded, and stored in a vector database. During chatbot conversations, user queries are embedded and matched against the knowledge base using semantic similarity search, with retrieved documents injected into the LLM prompt as context. This retrieval-augmented generation (RAG) approach grounds chatbot responses in client-specific information rather than relying solely on the base LLM's training data.

Enables deployment of the same chatbot logic across multiple communication channels — web widget, SMS, WhatsApp, Slack, Teams, or voice (phone/IVR). The platform likely uses a channel abstraction layer that translates between different message formats and APIs while maintaining consistent conversation state and context across channels. Each channel integration handles protocol-specific requirements (character limits for SMS, rich formatting for Slack, audio transcription for voice).

Provides real-time and historical analytics on chatbot conversations, including intent recognition accuracy, user satisfaction metrics, conversation drop-off points, and response latency. The dashboard likely tracks metrics like conversation completion rate, average session duration, top intents, and user feedback (thumbs up/down). Agencies can drill down into individual conversations to debug failures or identify training opportunities for the chatbot.

Allows chatbots to call external APIs or trigger custom actions during conversations — e.g., look up customer account status, create support tickets, process payments, or fetch real-time data. The platform likely provides a node-based interface for defining API calls (HTTP method, headers, payload) with variable substitution from conversation context. Responses from APIs are parsed and injected back into the conversation flow or used to conditionally branch the chatbot logic.

Automatically classifies user messages into predefined intents (e.g., 'billing question', 'product inquiry', 'complaint') and routes conversations to appropriate responses or escalation paths. The platform likely uses the underlying LLM to perform intent classification with few-shot examples, or a lightweight classifier trained on agency-provided intent definitions. Confidence scores determine whether to execute the matched intent or escalate to human support.

Enables seamless escalation from chatbot to human agents when conversations exceed chatbot capabilities or confidence thresholds. The platform likely maintains conversation context (history, intent, customer info) and routes to available agents via email, Slack, or integrated ticketing systems. Agencies can define escalation rules (e.g., 'if confidence < 0.7, escalate to human') and track handoff metrics (escalation rate, time to human response).

Maintains conversation state across multiple turns, allowing the chatbot to reference previous messages and build context for coherent multi-turn interactions. The platform likely stores conversation history in a database with per-user session management, and injects relevant history into the LLM prompt to maintain context. Agencies can configure context window size (how many previous messages to include) and conversation timeout (when to start a new session).

Allows agencies to define response templates with variable substitution and conditional branching based on conversation context, user attributes, or API responses. Templates likely support Handlebars or Jinja2-style syntax for injecting variables (e.g., 'Hello {{customer_name}}, your balance is {{account_balance}}'). Conditional logic enables different responses based on intent confidence, user segment, or API response data (e.g., 'if inventory > 0, say product is available; else, offer backorder').