Voice Agent Platform
A white-label, multi-tenant AI agent platform for enterprise voice and chat — retrieval-backed knowledge, tool use, and compliance-ready flows. Agents that act, not chatbots that reply.
At a glance: Ethereal built an AI phone-and-chat assistant that answers customer calls and messages around the clock, understands what people actually need, and gets things done instead of just reading canned replies. It replies in under a second, works for both phone and web chat, and follows US privacy and calling rules automatically.
01The Problem
Businesses — including US-regulated ones — need 24/7 voice and chat support, but face high telephony costs, strict compliance exposure (TCPA, PCI, PII), and slow per-client onboarding. Most "AI support" products are chatbots that only reply; enterprises need agents that act.
Round-the-clock human phone coverage is the single most expensive support line item.
TCPA, PCI, and PII rules make every automated call a legal exposure if handled naively.
Onboarding each new client meant weeks of custom engineering — it had to become config.
The Solution
We designed a white-label, multi-tenant platform on a 4-layer architecture — tenant config → stack profile → industry pack → engine — so new tenants onboard with zero code. Each tenant gets its own FAISS-backed retrieval knowledge base. Compliance gates are on by default: consent matrix, automatic PII redaction, PCI-pause during card numbers, DNC registry checks, and explicit AI disclosure. A single engine serves both telephony (voice) and WebSocket (chat) channels.
Tenant Config
Brand, prompts, escalation rules, and knowledge sources — declared, not coded.
Stack Profile
Model, STT/TTS voices, and channel mix selected per tenant from vetted profiles.
Industry Pack
Vertical-specific flows, vocabulary, and compliance presets for regulated domains.
Engine
One agent runtime — retrieval, tool use, and conversation state — shared by every tenant.
End-to-end, per voice turn
The same engine answers web chat over WebSocket — same knowledge base, same tools, same compliance gates. One brain, two mouths.
Postgres holds tenant state; Redis keeps conversation memory hot enough to stay inside the latency budget.
Every outbound contact is checked against recorded consent before a call is placed.
Personal data is redacted automatically from transcripts, logs, and model context.
Recording and transcription pause the moment a caller reads out card numbers.
Do-Not-Call registry checks run before any outbound campaign touches a number.
Callers are explicitly told they are speaking with an AI agent — always, by default.
The Outcome
A ~750ms end-to-end voice latency budget keeps conversations human-grade. One unified engine serves two channels, and compliance is built in for regulated verticals — new tenants onboard without touching code.
Built with
08 technologiesHave an agentic ai problem like this?
Tell us what you're building — we'll tell you exactly how we'd ship it.
From the archive
An AI sales agent that moves leads from first message to qualified opportunity — reading intent, pricing dynamically, and alerting a human the moment a lead turns hot.
An AI home-advisor for real estate that qualifies buyers, scores its own confidence, and escalates hot leads over WhatsApp — tuned by a three-layer evaluation suite.
Three parallel AI agents that audit websites, product flows, and user experiences from multiple perspectives at once — with video evidence.
Describe a multi-leg options strategy in plain English — the copilot constructs it, validates it against risk rules, and executes across accounts.