HeyGen Avatar Concierge
A persistent video-avatar agent that greets, qualifies, and onboards every visitor face-to-face
HeyGen Avatar Concierge wraps the HeyGen v3 streaming-avatar API behind a Composer subagent. It runs scripted and dynamic conversations — sales discovery, onboarding walkthroughs, support triage — as a real-time avatar with the customer's brand voice, then hands the transcript to Composer for follow-up.
Overview
Concierge is the customer-experience surface of Hannah Imagineer. Every conversation is a structured turn-by-turn run that Composer can replay, branch, or hand off to a human. Avatar identity, voice, and script are versioned per campaign; the underlying HeyGen integration lives in lib/heygen/.
Primitives
Real-time streaming video avatar (HeyGen v3)
Script-driven and dynamic conversational modes
Lead qualification with structured field extraction
Onboarding walkthroughs synced to product state
Transcript hand-off to Composer for follow-up plans
Per-campaign avatar identity and voice versioning
Outcomes
- 01Replace the empty homepage hero with a face-to-face greeting
- 02Qualify leads before a human ever picks up the conversation
- 03Onboard new customers without scheduling a live call
Integrations
Autonomy & guardrails
- Outbound contact actions (email, SMS) outside the configured campaign
- Sharing transcripts outside the configured CRM destination
Guardrails & requirements
Guardrails
- Conversations bounded by a per-campaign script + intent
- PII captured during qualification is routed through the Privacy Router
- Avatar usage rights enforced per campaign
Requirements
- HeyGen API key (v3)
- A configured Composer workspace
Technical specifications
Runtime
- Harness
- HeyGen v3 streaming + Composer subagent
- Deployment
- Managed via Composer workspace
- Data residency
- Customer Supabase + HeyGen-hosted avatar render
- License
- Commercial · Hannah Imagineer EULA · HeyGen subprocessor
- Version
- v1.0
Models & tooling
- Models
- claude-sonnet-4-6heygen-avatar-v3
- Tooling
- heygen-rest-wrappertranscript-extractorintent-classifier
Reliability & limits
- Latency SLA
- ~700ms p50 first-token avatar latency
- Rate limits
- Bounded by HeyGen account concurrency
Security & compliance
- Auth model
- Composer SSO · HeyGen API key per workspace
- Compliance
- SOC 2 Type II (in process)GDPR
Evals & benchmarks
No public benchmark applies cleanly to a streaming-avatar concierge. The internal eval set scores intent-completion, qualification-field accuracy, and first-token latency on a held-out conversation set; numbers will publish alongside the Composer-native benchmark release.
Architecture notes
Memory architecture
Per-conversation transcripts are persisted to the customer Supabase. Long-lived customer context is read from Shared Mind by tenant; per-conversation scratchpad is garbage-collected after hand-off.
Context strategy
The conversation is shaped by a campaign-level script (system prompt + intent tree) plus turn-by-turn retrieval from Shared Mind when the avatar needs product context.
Evaluation
Evaluation runs against a held-out conversation set scored on three axes: intent-completion rate, qualification-field accuracy (vs. human label), and avatar latency. Live runs are sampled into the next eval cycle.