Platform
One assistant,
six connected parts
Orrin is the knowledge index, the action catalog, the identity model, the safety layer, the evaluation suite and the analytics that sit behind one command bar in your product. Everything on this page describes the product being built with design partners.
Knowledge
Answers from
your documentation
Connect a public documentation or help center URL, or upload Markdown, HTML, text, JSON and CSV. PDF arrives with design partners. Documents are chunked, embedded and stored in a search index that belongs to your workspace alone.
What it does
- Scheduled resync every day, plus a Resync button
- A list of documents with page counts and ingestion status
- Deleting a document removes its chunks within minutes
- Answers carry numbered citations that open the source document
Actions
Tools from
your OpenAPI file
Import an OpenAPI specification and Orrin converts endpoints into candidate tools with names, descriptions and parameters. You choose which endpoints are exposed and label each one Read, Write or Sensitive.
What it does
- Nine property management templates ready to map to your endpoints
- Sensitive actions require a typed confirmation from the end user
- Optional: expose the catalog as a hosted MCP server for other agent surfaces
- No bulk destructive actions, evictions or screening decisions in the first release
Identity
Every call
as the signed in user
Your backend mints a short lived signed token for each end user, carrying user id, company id, role and scopes. The widget passes it with every request, and Orrin calls your API as that user.
What it does
- Tokens live between five and fifteen minutes
- Never more permission than the user already has
- No shared service account for end user actions
- Every tenant isolated at every layer: index, logs, audit log, credentials
Safety
A preview
before every change
Before any Write action the assistant shows what will change, the exact values and which record. The user taps Confirm or Cancel. Sensitive actions need a typed word. Undo is offered where your API supports reversal.
What it does
- Rate limits per user and per company
- Fair housing guardrails and refusals for steering, legal advice and screening
- Retrieved documents and API responses are treated as data, never as instructions
- A complete audit log per action: user, company, tool, parameters, preview, confirmation, response, time
Evaluation
A gate
in front of production
Each workspace has a test suite of real questions and action scenarios, and the property templates come with a starter suite. Staging and Production exist from day one, and nothing reaches Production without passing.
What it does
- Scores: grounded answer rate, citation accuracy, correct tool choice, correct parameters, confirmation compliance, correct refusals
- Publishing is blocked below one hundred percent confirmation compliance or on any failed policy test
- Versions compared side by side before publishing
Analytics
Usage you can
pass through
Conversations, resolution rate, actions completed, top questions and unanswered questions. Usage is metered per vendor and per end customer company so you can add AI costs to your own invoices.
What it does
- Conversations, model tokens, retrieval calls and actions executed per company
- Usage export and a webhook
- Personal data redacted from analytics views; full content stays in your isolated log
Workspace
Two environments,
three roles
Sign in
Email and password, plus Google and Microsoft sign in.
Roles
Owner, Developer and Support lead, each with what they need and nothing more.
Environments
Staging and Production from day one. Nothing reaches Production without passing its evaluation suite.
The embeddable assistant
A command bar
that wears your brand
Two ways in
A script tag plus one initialization call, or the npm package for React apps. Opens on Cmd K or Ctrl K, or from a launcher button.
Plain language states
Answers with numbered citations, action preview cards, confirmation, success and failure, all in words a property manager uses.
History and handoff
Conversation history per user, and a handoff action that creates a support ticket with the conversation attached.
Themeable
Colors, font, corner radius and logo. The default theme is the cream panel with maroon text and a mustard Confirm you see in the demo.
Latency targets
First token under one and a half seconds, a grounded answer under four seconds, an action preview under six seconds. These are targets for the first release, not measurements, and they will be published as measurements with the method once the assistant is live.
Coming later
Listed as planned,
not built
- Voice mode through the microphone in the command bar
- Proactive suggestions, such as noticing an overdue work order while a unit is open
- A second vertical: logistics and transport management software
- Slack and Microsoft Teams surfaces for the same assistant
- Multilingual answers
- SOC 2 audit