Platform Docs

Welcome to Chatter Genius

Chatter Genius is the orchestration layer for the modern AI workforce. We enable organizations to configure intelligent AI Agents for individual voice, SMS, or email workflows. The platform combines managed communication delivery, AI inference, optional knowledge retrieval, configurable safety controls, and post-conversation analytics in one operating console.

The orchestration layer

Use this guide to plan, launch, and scale AI Agents. You will learn how to select role-based templates, optionally equip eligible agents with knowledge bases, configure guardrails, schedule a workflow on one channel, and interpret the message- and conversation-level signals produced after an interaction completes.

What you can do today

• Configure AI Agents — Start from reusable role templates for customer support, sales development, HR, IT help desk, scheduling, onboarding, and other guided-conversation use cases, or create a custom agent.
• Run channel-specific workflows — Create, schedule, launch, monitor, and archive an individual outbound voice call, two-way SMS thread, or threaded email conversation. Each workflow uses one channel and one destination.
• Ground eligible agents in your documents — Pro, Business, and Enterprise plans can upload supported documents to per-agent knowledge bases. The platform chunks and embeds content, then retrieves relevant context when knowledge retrieval is enabled and succeeds.
• Apply conversation controls — Configure per-agent allowed and forbidden topics, regular-expression blocks, toxicity thresholds, escalation behavior, and maximum turns. Platform-owned sensitive-value rules redact detected high-confidence PII and MNPI patterns before text persistence and onward egress.
• Review post-conversation intelligence — Completed conversations can produce sentiment, emotion, intent, frustration, estimated CSAT, agent-quality, compliance, language-quality, topic, outcome, and summary signals. Individual enrichments can abstain or degrade when confidence or a dependency is unavailable.
• Choose a supported AI configuration — Platform-managed AI currently uses OpenAI. Bring-your-own-LLM text configurations support OpenAI, Azure OpenAI, Google Gemini, Anthropic, and custom OpenAI-compatible endpoints; realtime voice supports OpenAI and Azure OpenAI. AWS Bedrock configuration is coming soon and is not currently credential-validated.
• Track capacity and billing — Basic, Pro, Business, and Enterprise plans include stated call, voice-minute, and outbound SMS and email initiation allowances. The billing service records initiated calls, completed call duration rounded up to whole minutes, and successful initial outbound SMS and email workflow sends for supported usage and overage calculations. Later conversation replies do not currently add SMS or email usage units. Monthly subscriptions use monthly usage periods; under the current Early Access billing implementation, an annual subscription receives the stated allowances once for its annual usage period rather than a monthly reset.

How it works

1. Request access: Early Access registration is limited to approved participants. Once invited, register and verify your email and phone by one-time passcode.
2. Subscribe: Choose a capacity tier (Basic, Pro, Business, or Enterprise) and select platform-managed AI or a supported bring-your-own-LLM configuration. Twilio and SendGrid communication delivery remains platform-managed in both modes.
3. Configure: Create your first AI Agent from scratch or from an available role template. Define its name, role, objectives, language hint, tone, instructions, topic controls, and turn limit. Pro and higher plans may also add a per-agent knowledge base.
4. Launch: After the communication profile finishes provisioning, create a voice, SMS, or email workflow for one destination. Schedule it for a future time or start it immediately.
5. Monitor: Track workflow status through its full lifecycle: created, scheduled, running, analyzing, completed, analyzed, failed, canceled, or archived. The dashboard reports recorded conversation and subscription-usage data; it is not a streaming analytics feed.
6. Analyze & Optimize: After each conversation completes, the analytics service processes its transcript asynchronously and can generate message-level analytics (sentiment, emotion, intent, frustration, language quality, topics, compliance flags) and conversation-level summaries (TL;DR, pros and cons, actionable insights, agent quality scores across empathy, relevance, completeness, professionalism, and resolution effort, CSAT, engagement metrics, and outcome classification). Health metadata identifies defaulted or degraded enrichments.
7. Scale: Provision additional agents, expand your knowledge bases, and upgrade your capacity tier as demand grows.

Communication channels

• Voice — Outbound calls via Twilio with real-time AI conversation over WebSocket audio streaming. The managed configuration uses the OpenAI Realtime API; supported customer-supplied realtime configurations use OpenAI or Azure OpenAI. The realtime provider receives audio before the platform receives a final transcript for local filtering.
• SMS — Outbound and inbound two-way SMS via Twilio. Conversations are threaded by phone number pairs, enabling persistent context across message exchanges.
• Email — Outbound and inbound email via SendGrid. Conversations are threaded using custom headers and reply-to encoding, preserving full dialogue history across exchanges.

Channel history is maintained within its workflow. Separate voice, SMS, and email workflows can reuse the same agent configuration, but the platform does not currently stitch conversation history across channel transitions.

Knowledge base

Knowledge bases are available on Pro, Business, and Enterprise plans. Basic does not include knowledge-base access.

• Supported formats — PDF, Microsoft Word (.docx), PowerPoint (.pptx), Excel (.xlsx), CSV, JSON, HTML, Markdown, and plain text. Plain-text formats are parsed locally; rich documents are sent to the LlamaParse cloud service for text extraction when that service is configured.
• Semantic retrieval — Documents are automatically chunked and embedded into a pgvector-backed vector store. When retrieval is enabled and relevant chunks are available, the platform injects bounded context into the LLM prompt.
• Scoped storage and retrieval — Storage keys and vector queries are scoped by tenant and agent, and vector rows are protected by Postgres row-level security. Upload limits (total storage, document count, file size, and PDF page count) are governed by the subscribed tier.
• Optional strict fallback — A deployment setting can require a predefined fallback message when relevant context is unavailable. This setting is disabled by default and is not currently a per-agent control.

Guardrails engine

Text and final voice transcripts pass through layered controls. A local, non-overridable sensitive-value boundary scans user and assistant text before text persistence or onward service egress. Other policy layers depend on configuration and provider availability, and the realtime voice provider processes audio before a final transcript reaches this boundary.

1. Small-talk bypass — Greetings and simple acknowledgements are recognized and handled without triggering heavy moderation, preserving low-latency responsiveness.
2. Deterministic policy rules — Python-owned rules evaluate configured topics, regular expressions, repetition, and other conversation policy conditions. Matching rules may block, replace, or escalate a response according to the active configuration.
3. Semantic topic classifier — Configured embeddings measure semantic similarity between conversation content and approved or blocked topic definitions when the embeddings dependency is available.
4. Content moderation — The configured moderation provider can screen for categories such as hate, harassment, self-harm, violence, and sexual content. Availability and fail-open/fail-closed behavior are deployment settings.
5. Sensitive-value policy — Platform-owned YAML definitions detect configured high-confidence PII, credential, and MNPI patterns and replace matched values with explicit redaction markers. The controls reduce exposure but cannot guarantee detection of every sensitive value.
6. Unsupported-action protection — The conversation runtime blocks claims that it completed unsupported external actions such as refunds, bookings, account changes, transfers, or ticket creation.

Conversation safety gate

In addition to content guardrails, the platform enforces conversation-level safety policies:

• User disengagement detection — Over 120 phrases across 9 languages are recognized as disengagement signals, with LLM fallback for ambiguous cases.
• Opt-out assistance — Common opt-out signals such as STOP and UNSUBSCRIBE are detected, persisted to a suppression list, and checked before outbound SMS and email. Customers remain responsible for their legal obligations.
• Off-topic streak enforcement — Consecutive off-topic turns trigger automatic conversation closure to prevent abuse.
• Maximum turn limits — Configurable per-agent turn caps prevent runaway conversations.

Post-conversation analytics

After a completed conversation is handed off, the analytics service asynchronously attempts message-level and conversation-level enrichment. Timeouts, low-confidence classifications, and unavailable dependencies can produce abstentions or bounded defaults, which are recorded in analytics health metadata.

• Sentiment analysis — Per-message multilingual three-class sentiment with an English-only VADER fallback, aggregated into overall scores, per-speaker breakdowns, and trajectory trends.
• Emotion classification — Multilingual zero-shot detection across the canonical five-label ontology (anger, fear, sadness, joy, neutral), with dominant emotion identification and per-speaker distributions.
• Intent detection — User messages are matched against the maintained intent catalog, with a sentence-transformer fallback and explicit unknown outcomes when confidence is insufficient.
• Frustration scoring — Multi-signal blend (sentiment 40%, emotion 40%, text cues 20%) with escalation detection when thresholds are exceeded.
• Customer Satisfaction (CSAT) — Multi-signal 1–5 scoring that synthesizes sentiment, resolution status, and engagement patterns.
• Agent quality assessment — LLM-scored across five dimensions: empathy, relevance, completeness, professionalism, and resolution effort (1–5 scale).
• LLM-assisted summaries — TL;DR summaries, strengths-and-weaknesses analysis, and actionable business recommendations are requested after a conversation. Failed or structurally invalid provider responses are marked as degraded instead of being presented as complete results.
• Post-conversation compliance signals — Keyword matching, regular expressions, and NER-based detection flag potential issues at the message level. These analytics are distinct from realtime guardrail interventions.
• Language quality — Grammar, spelling, and style checks via LanguageTool with readability scoring.
• Topic extraction — KeyBERT-powered keyphrase extraction identifies the dominant topics in each conversation.

Full language-specific analytics routing currently covers English, Spanish, French, Polish, and Russian. Some classifiers use cross-lingual or zero-shot behavior, and their output should be treated as an estimate rather than a definitive assessment.

Platform strengths

• Reusable agent configuration: Define objectives, tone, language, scripts, topic controls, and a turn limit once, then use that agent in separate voice, SMS, or email workflows.
• Layered safety controls: Sensitive-value redaction, deterministic rules, semantic classification, moderation, suppression checks, and bounded conversation behavior work together according to the active deployment configuration.
• Channel-specific continuity: Two-way SMS and email maintain history within their own threads. Separate channel workflows do not currently share a unified conversation history.
• Interaction intelligence: Post-conversation analysis combines message signals, outcome indicators, estimated CSAT, agent-quality scores, and LLM-assisted recommendations, with degradation metadata for incomplete results.
• Predictable Pricing: Tiered subscription plans (Basic, Pro, Business, Enterprise) combine a base subscription with included allowances and published overages for initiated calls, rounded-up completed-call minutes, and successful initial outbound SMS and email workflow sends.
• AI-provider flexibility: Use the platform-managed OpenAI configuration or connect a supported customer-supplied LLM account. Twilio and SendGrid remain managed platform communication providers.

Security and access

• RSA-Signed Authentication: RS256 JWT access and refresh tokens with rotation, revocation, and Redis-backed blacklisting.
• Role and ownership checks: User, administrator, and internal-system roles are combined with owner-or-administrator checks on user-facing resources. The current product does not provide organization membership or custom-role administration.
• Login Protection: Per-email and per-IP lockout, rate limiting, OTP attempt limits, and optional carrier/VoIP screening help reduce brute-force and abuse risk.
• Transport and secret handling: External platform ingress requires TLS 1.2 or later. Platform service credentials are supplied through Google Cloud Secret Manager; customer-supplied LLM keys are encrypted in the application database.
• Service boundaries: Internal service operations use service credentials with optional previous-key windows to support rotation, while Twilio, SendGrid, and Stripe callbacks use provider signatures or configured verification tokens.
• Conversation controls: Sensitive-value detection, moderation, topic controls, suppression checks, and circuit breakers reduce risk but do not guarantee that every unsafe input or output will be detected.
• Operational metadata: Conversation records can include redaction, guardrail, delivery, and analytics metadata. This is not a general-purpose compliance audit-log product.

Integrations and API access

Chatter Genius does not currently offer packaged CRM or HR connectors, a public developer API-key program, or customer-configurable outbound webhooks. The product console uses authenticated application endpoints; Twilio, SendGrid, analytics, and billing service interfaces are internal platform boundaries rather than a public integration API. Contact us to discuss integration requirements.

Getting started

1. Request Early Access: Contact us for an invitation. Approved participants can register and verify email and phone by one-time passcode.
2. Choose a plan: Select a capacity tier that fits your scale. Plans range from Basic through Enterprise, with current prices and allowances listed on the Pricing.
3. Provision your Agent: Create an agent from scratch or start with a role template. Define the agent's name, role, language, instructions, objectives, and topic controls.
4. Build a knowledge base when eligible: Pro and higher plans can upload supported documents for chunking, embedding, and semantic retrieval.
5. Set guardrails: Configure blocked topics, PII and MNPI detection profiles, toxicity thresholds, escalation triggers, and maximum turn limits.
6. Create and launch a workflow: Define the workflow type (voice, SMS, or email), assign an agent, specify origin and destination contacts, and schedule or start immediately.
7. Review & Improve: Use the Workflow Insights dashboard to examine conversation transcripts, review available post-conversation signals, and refine agent instructions based on actual interaction data.

Need help?

Questions or feedback? Our technical team is here to help.

• Contact us with product or technical questions.
• Already a customer? Reply to any support email or reach out at support@chattergenius.ai.
• Evaluating Chatter Genius? Request a technical deep-dive via our contact form.