Which AI models are supported?

Delveira is designed around provider interfaces. The first implementation uses OpenAI for listening and research planning, with room for additional models. Model choice can vary by task—transcription, claim structuring, and research routing do not have to share a single model.

How is session and transcript data handled?

Session state, transcripts, claims, and evidence are persisted in your application database. Provider API keys never ship to the browser. Exact retention and regional controls for hosted SaaS will be documented when authentication and organizations ship.

How do citations and provenance work?

Every displayed factual item is intended to preserve provenance from the UI card through evidence passages to the source. Sources keep URL, title, publisher or author when available, retrieval time, and content type. Absence of evidence is never represented as “false.”

When will subscriptions be available?

Plan tiers and pricing on this site are live and current. Self-serve checkout is not open yet—join the waitlist and we will reach out as paid tiers open.

How will billing work?

Billing integration (invoices, payment methods, and self-serve upgrades) is coming in a later phase. Until then, join the waitlist and we will help you get set up directly.

Realtime research for presenters

Structure claims. Gather evidence. Present with confidence.

Delveira listens to a live discussion, turns speech into revisioned claims, gathers provenance-aware evidence, and helps you publish a clean audience view—without treating AI output as unverified truth.

Pull the thread. Follow the evidence.

Join waitlist Explore features

Already have access? Launch App →

What it is

A research control room—not a chat window

Every claim you make stays structured, revisioned, and traceable. Nothing gets presented to your audience without you reviewing it first.

Listen live
Explicit microphone control, partial transcript deltas, and durable finalized segments.
Inspect evidence
Supporting, qualifying, challenging, or absent—never confuse silence with falsehood.
Publish carefully
Compose cards in the control room, then project a recording-friendly presentation route.

How it works

Four steps from speech to slide

Built for a single presenter operating one session—with a clear path to richer collaboration later.

01
Listen
Capture live transcription with explicit start/stop control. Partial text appears quickly; finalized segments drive analysis.
02
Structure
Convert speech into revisioned claims, topics, and entities—typed domain objects, not freeform chat blobs.
03
Research
Run seeded and public research in parallel. Classify evidence as supporting, qualifying, challenging, or absent.
04
Present
Compose cards in the control room, then publish a clean audience view suitable for recording.

Features

Capability highlights

Concise building blocks for live inference, research, and presentation.

Live inference

Turn finalized transcript segments into structured topics, claims, and research commands as the conversation unfolds.

Multi-model AI

Route listening, planning, and research through provider interfaces so models can specialize by task.

Evidence gathering

Retrieve passages—not just summaries—from seeded documents and the public web, with clear stance labels.

Research routing

Generate support- and challenge-oriented queries, then send work to the right source families.

Government archives

Reach declassified FOI and diplomatic records through specialized archive providers when relevant.

Academic sources

Connect scholarly indexes and library catalogs so claims can be grounded in peer-reviewed material.

See all capabilities

Research

Research capabilities

Seeded corpora, public web, and specialized providers—routed by claim, not by guesswork.

Seeded corpora

Search uploaded Markdown and PDF material with hybrid full-text and vector retrieval.

Public web

Plan queries that seek both support and challenge, then extract citable passages.

Specialized providers

Route to archives, scholarly indexes, and ancient-text corpora when the claim warrants it.

Under the hood

The right model for every step

Listening, claim structuring, research planning, and embeddings each use a model suited to the task—so quality keeps improving without you changing how you work.

Live transcription
Claim structuring
Research planning
Semantic embeddings

Built to improve underneath you

Models can be upgraded or swapped per task without disrupting your sessions or changing your workflow. Your credentials and API secrets never leave the server.

Structured, not freeform

Every model response is validated into claims, evidence, and cards—never raw, unchecked output rendered straight to your audience.

Evidence & citations

Provenance from card to source

Every factual item should remain inspectable. Citation metadata travels with the evidence.

Evidence passages keep source URL, title, publisher or author when available, retrieval time, and content type. Stance labels help you see support and challenge side by side.

SupportingPassages that corroborate the claim
QualifyingNuance that narrows or conditions the claim
ChallengingMaterial that conflicts or complicates

Live inference

Keep pace with the conversation

As claims revise, obsolete research is superseded so stale results do not overwrite newer understanding.

Partial transcript text appears quickly for awareness. Finalized segments drive semantic listening: topics, entities, claims, comparisons, questions, and presentation commands.

Learn about live inference

Declassified archives when the claim warrants it

Specialized providers can reach FOI and diplomatic primary records—without putting archive credentials in the browser.

History Lab and beyond

Declassified archive research is available through dedicated integrations. Enable or disable them per environment; the product remains useful without them.

Ancient texts, scholarly indexes, and library catalogs follow the same pattern: provider behind an interface, provenance on the result.

Pricing