How It Works

From governed knowledge to real-time, multi-channel AI experiences.

Runtime loop

A predictable interaction cycle that stays fast, grounded, and consistently on brand.

Signal path for every interaction. Each stage is observable and governed.

Kiosk greeting scene

Detect & greet

Channel-aware entry that tailors the first response to the moment.

  • Proactive triggers for kiosk, web, and call center
  • Language detection plus greeting templates
  • Configurable prompts and entry rules in the platform
Speech input and transcription

Listen & understand

Streaming speech and text intake with optional vision.

  • Real-time transcription for fast routing
  • Web Speech and telephony audio input
  • Context capture for downstream retrieval
Emotion recognition from voice, video, and text

Sense emotion & respond

Real-time emotional cues help the assistant stay empathetic, calm, and on-brand.

  • Emotion recognition from voice, video, and text in real time
  • Stabilized emotion state with confidence scoring (avoids "jittery" reactions)
  • Drives tone, pacing, escalation triggers, and avatar facial expressions
Live interruption handling

Interrupt & recover

True barge-in with playback cancel and late chunk suppression.

  • Stop playback immediately on new intent
  • Cancel active run and ignore late audio
  • Resume in context without re-asking
Knowledge retrieval interface

Retrieve & ground

Hybrid retrieval keeps answers anchored to approved knowledge.

  • Vector and keyword search with reranking
  • Multilingual embeddings for consistent coverage
  • Source-aware responses for safer answers
Flow decision routing

HoloBrain

Realtime agent orcestrator

  • Channel-specific rules and tool gating
  • Multi-agent quick response while tools run
  • Escalation policies when confidence is low
  • Smart data managment from api's and vector databases
  • Realtime agent, flow and tool caling
3D avatar facial animation

3D avatars & facial animation

HDRP + advanced face rig for production-grade 3D avatars for live dialogue.

  • Audio-driven, language-independent lip sync
  • 60+ blend shapes with emotion and eye motion
  • Photoreal materials for skin, eyes, and hair
Workflow actions in progress

Act & guide

On-screen actions and workflows move the task forward.

  • UI cards, forms, and webviews triggered by flows
  • Tool execution with status updates
  • Unified behavior across pipelines
Human handoff and coaching

Escalate & learn

Human handoff with full context and analytics feedback.

  • Transcript and metadata passed to operators
  • Outcome logging for continuous improvement
  • Per-client reporting and health checks

Knowledge ops

From raw content to optimized, governed knowledge that shortens time to value and boosts answer quality.

FastAPI ingestion request and response
FastAPI intake

Unified knowledge intake

One gateway for every format, so onboarding is faster and governance stays tight.

  • Q and A, PDFs, HTML, and media in a single schema
  • Auto-validation and dedupe to keep the corpus clean
  • Agent and channel tags for brand-safe routing
PDF input
OCR result
OCR result
PDF input
OCR pipeline

OCR to structured knowledge

Scanned PDFs become searchable, cited knowledge so nothing stays trapped in images.

  • Detect and pull PDFs from sitemaps or folders
  • OCR to clean Markdown with headings and tables
  • Source URLs and timestamps for auditability
Q and A generation output
Q and A generation

Q and A at scale

Generate consistent, on-brand answers in hours instead of weeks.

  • LLM-assisted Q and A from document chunks
  • Coverage for common questions and edge cases
  • Reviewable JSON output for rapid QA
Synthetic data generation preview
Synthetic data

Synthetic data enrichment

Fill gaps and stress-test the assistant before real customers do.

  • Simulate long-tail intents and rare scenarios
  • Balance training data across topics
  • Boost routing accuracy and fallback confidence
LLM knowledge optimization output
LLM optimization

LLM knowledge optimization

Transform raw content into concise, reliable knowledge that models can trust.

  • Normalize terminology and remove duplicates
  • Summaries, keywords, and topic anchors for retrieval
  • Tone alignment per client and channel
Retrieval optimization report
Retrieval tuning

Retrieval optimization

Higher precision, lower latency, and fewer hallucinations at scale.

  • Hybrid search tuning with reranking
  • Multilingual embeddings aligned across sources
  • Relevance metrics to tune and prove ROI
Automatic knowledge agents generation output
Agent generation

Automatic agent generation

Specialist agents built from your knowledge for faster, safer handoffs.

  • Cluster domain knowledge into dedicated agents
  • Auto-built prompts, keywords, and handoff rules
  • Scale to new business units without rework

AI flows & pipelines

A single real-time pipeline from capture to handoff powers kiosk, web, voice, 3D, and robotics.

01

Capture and validation

Speech input capture

Turn voice or UI events into a clean, governed request in real time.

  • Streaming STT for kiosk, web, and SIP
  • Language detection plus session metadata (channel, locale, device)
  • Optional presence signals for greeting and input checks
Configurable via: realtime-v1 user_turn inputHints (modality, locale) + variables node (session metadata)
02

Knowledge ingestion & DataAPI retrieval

Knowledge retrieval interface

Keep answers grounded in approved content and live data sources.

  • Ingest docs, FAQs, and media into vector indexes with agent tags
  • Retrieve context with embeddings and structured search, then rank and inject only the best snippets
  • DataAPI tools add live sources with schema-validated calls and domain prompt injection
Configurable via: document loaders/splitters, vector store + embedding + record manager config, DataAPI tool registry (DataApiCallable) + APIPrompt.md
03

Orchestrator (policy-aware)

Orchestrator routing view

Decide what happens next under guardrails.

  • Build the response envelope with persona, policies, channel rules, and summaries
  • Route to the best model or provider for latency vs quality
  • Choose answer, tool call, Flowise flow, or specialist agent
Configurable via: realtime-v1 agent node (system_prompt, policies, guardrails, memory, variableRefs)
04

AgentFlow and tool selection

AgentFlow routing canvas

When work needs steps, move from chat to structured execution.

  • Flow selection uses titles, descriptions, and required schemas
  • Tool registry and allowlists restrict actions
  • Specialist agents handle deep data collection while the main AI stays in control
Configurable via: agent.tools (name/kind/config/safety) + Flowise flow JSON + exec schemas
05

Execution and integrations

Tool and action execution

Turn decisions into outcomes across systems and on-screen experiences.

  • Call APIs and connectors (CRM, booking, internal systems)
  • Trigger runtime commands like panels, QR, forms, or media capture
  • Log actions per turn for audit and replay
Configurable via: action schemas + tool credentials (headers, keys)
06

Streaming delivery and handoff

Streaming voice delivery and handoff

Respond immediately, stream naturally, and escalate without losing context.

  • Partial response streaming for low-latency feedback
  • TTS streaming with interruption and per-turn voice locking
  • Channel adapters for kiosk, web, render streaming, robotics, and SIP with human handoff context
Configurable via: agent voice/persona hints + TTS settings + realtime-v1 export

Platform

Visual builders, integrations, and controls for production-grade agents.

Agentflow V2 + Realtime-v1

Visual builders

Design single- and multi-agent workflows plus realtime graphs.

  • Branching, routing, and tool gating in a visual editor
  • Realtime-v1 outputs compact JSON for Unity/runtime clients
  • Render Streaming UI sync for chat, webviews, and cards/offers
  • Multi-user sessions with keepalives and chat-only fallback
Prompting + templates

Prompt governance

Draft, publish, and reuse prompts across channels.

  • Chat, few-shot, and structured prompt templates
  • Draft vs published prompts for safe rollout
  • Input moderation and output post-processing controls
Tools + MCP

Custom tools & integrations

Connect APIs, actions, and external systems into flows.

  • HTTP, OpenAPI, and custom tool nodes
  • MCP client/server nodes with auth support
  • Allowlisted browsing and action hooks
Data ingestion & integration

Vector databases & RAG

Connect sources, build indexes, and power retrieval.

  • FastAPI ingest for Q&A/media with dry-run, schema auto-create, and agent tags
  • Hybrid vector + keyword retrieval with tunable RRF/top-k
  • Curated media plus optional face recognition with blur/consent checks
Security controls

Credentials & secrets

Encrypted credentials and isolated configs per customer.

  • Encrypted credentials with per-instance keys
  • RBAC/SSO and restricted domains for access control
  • Per-client endpoints and key separation for isolation
Tracing + analytics

Execution monitoring

Trace runs, debug flows, and measure outcomes.

  • Execution logs and visual debugging
  • Per-client domains, ports, and analytics endpoints
  • Health checks via /_ping and /_ready plus count summaries
  • Offline caches and alerting for ops readiness
Channels + UX

Configurable speech & branding

Pick STT/TTS vendors and align UI to each brand.

  • Configurable STT/TTS vendors (Soniox, ElevenLabs, Azure)
  • Customizable embedded chat widget and components
  • Per-brand presets for logos, voice, language, and greetings

Trust & governance

Isolation, allowlists, and privacy guardrails for production AI.

Per-client isolation

Isolation by design

Separate domains, storage, and analytics per customer.

  • Scoped containers, ports, and analytics endpoints
  • Dedicated TURN and SIP configurations
  • Agent tags prevent cross-tenant bleed
Authenticated ingest

Controlled pipelines

Manifests and allowlists keep provenance clear.

  • Authenticated endpoints for ingestion
  • Allowlisted browsing and tool actions
  • Dry-run validation before full writes
Privacy + resilience

Policy-aware operations

Consent, retention, and reliability are enforced.

  • Opt-in camera and face features with consent checks
  • Per-client retention and access controls
  • Offline caches and prewarm steps reduce cold starts