Aura AI, under the hood.

What Aura AI is, in three paragraphs.

Aura AI is Sumeru’s AI Concierge — an enterprise-grade conversational presence layer that turns a business’s existing content into a face-driven, voice-led representative. It sells, supports, onboards, and interviews. It is, by design, far beyond a chatbot: an immersive conversational video experience that adapts in real time to who is on the other end.

It exists because customers and internal teams today lose 20–30% of their time navigating fragmented information — websites, PDFs, internal docs, disconnected tools. Aura collapses that into a single conversation. Powered by next-gen agentic AI, it instantly retrieves the most relevant answer from your knowledge and delivers it through a lifelike avatar — the same way a great human concierge would.

The platform is a six-layer stack — knowledge, conversation, voice, avatar, visual intelligence, and deployment — designed to clear a sub-second time-to-first-response budget and to render a 30 fps photoreal head on commodity GPUs. It is built for teams who need control, performance, and presence at scale: enterprises that demand SOC 2, GDPR, and HIPAA-aware data handling, and platform teams who want a self-hosted GPU option for sovereign deployments.

Six layers, one representative.

Each layer is independently scalable, observable, and replaceable. The orchestrator binds them with a single representative manifest.

Your content becomes the representative’s memory.

A representative is grounded by between one and eight sources. URLs are crawled to a depth you control. PDFs and Markdown are parsed with structure preserved — headings, lists, tables. Plain prompts are accepted for cases where the knowledge is the brand voice itself. Each source can be tagged for routing — a sales rep doesn’t read the engineering wiki by default.

Content is chunked at semantic boundaries, embedded into a 1024-dimension vector index, and attributed at retrieval time so every reply can carry a citation. Indexes refresh hourly by default, with on-demand reindex for time-sensitive launches. Pricing pages, inventory, and docs that change daily are first-class sources, not afterthoughts.

Two models, one conversation.

Every user turn first hits a routing pass on a 7B-parameter model. It classifies the turn — clarification, factual question, objection, action request — and selects the slice of the knowledge index to retrieve. A typical routing pass completes in under 80 ms.

The substantive reply runs on a 70B-parameter model with the retrieved context, the representative’s persona prompt, and the live conversation history. Output is streamed token-by-token to the voice layer so audio synthesis can begin before the reply finishes generating.

The orchestrator enforces representative scope. A sales rep cannot answer engineering questions; a support rep cannot quote sales discounts. Out-of-scope turns are gracefully redirected with a copy you control. Structured output — book a demo, capture a qualified lead — is emitted as a JSON tool call that downstream systems can consume.

The face — and the eyes — on top of the model.

Aura is a conversational video experience, not a chat thread with extra steps. Three distinct subsystems give it presence: a streaming voice that speaks in 28 languages, a photoreal avatar that lip-syncs to that voice, and a perception layer that watches the visitor in real time so the conversation can adapt to who is on the other end.

A second of presence, broken down.

Time-to-first-audio measured from the user’s last word. Targets are p95 on production traffic.

Five surfaces. One representative manifest.

A representative is described by a single manifest — its sources, persona, voice, avatar, and guardrails — and that manifest is the only thing that needs to travel between surfaces. Every surface listed below pulls from the same backend; switching surfaces does not require re-training, re-indexing, or re-uploading anything.

Built for teams that have to answer to a CISO.

Data is encrypted in transit (TLS 1.3) and at rest (AES-256). Conversation transcripts are tenant-isolated and retained only as long as your policy requires — Studio defaults to 30 days, Enterprise is configurable down to zero retention. Voice and avatar streams are never recorded by default. PII fields can be redacted before they reach the reply model. We are SOC 2 Type II audited; HIPAA-aware data handling and GDPR data residency in EU and US regions ship today; FedRAMP is on the 2026 roadmap.

Where the conversation goes.

Aura emits structured events for every conversation milestone — turn started, intent detected, lead captured, demo booked, escalation requested. Events fan out to your existing systems without polling. The Intelligent Agent Analytics dashboard gives you live engagement, conversion, and resolution metrics across every representative.

representative.session.started
representative.turn.completed
representative.intent.detected
representative.lead.captured
representative.demo.booked
representative.escalation.requested
representative.transcript.archived
representative.session.ended

What’s shipping, and what’s next.

A non-binding view of the next two quarters. Dates are intent, not contract.

Engineering questions, answered short.