TasiNexus: a cited, permission-aware assistant for staff and customers
Quantic Systems built TasiNexus, a multi-tenant assistant that answers employees and customers from a company's own documents. Every answer is cited and checked against its sources. Nobody sees a document they are not allowed to open.
- permission leaks across every test run
- 0
- answer correctness on the test set
- 99–100%
- test question matched to the right source
- Every
- prompt-injection attempts declined
- All
Measured on a graded test set, asked as each kind of user, with permission traps built in.
About Cedarline
Cedarline Equipment Rentals is the company we run on TasiNexus to show the platform end to end. It runs on production-grade cloud infrastructure, with a full set of company documents, staff accounts in separate teams, and customer conversations.
- Demonstration company
- Cedarline Equipment Rentals
- Knowledge base
- 70 documents, including scanned paper
- Users
- Employees in separate teams, a company admin, website visitors
- Channels
- Employee assistant, website chat, human support desk
- Cloud
- Microsoft Azure
- Status
- Live in a test environment. Business-system actions in progress
Three things keep company assistants from going live
Permissions
Many document assistants fetch everything, then try to hide restricted content afterward. Once a model has read a Finance memo, a prompt is the only thing between that memo and an intern.
Trust
An answer without a source cannot be checked. A customer bot that improvises a refund policy is a liability.
Handoff
When a bot cannot help, the customer is often sent to a new channel and has to explain everything again.
The Cedarline build had to solve all three. Staff ask about policies, incidents and pricing, and Finance content stays with Finance. Customers get answers from public documents only. A person can step in mid-chat and hand back without losing the thread.
What we built
One platform serves every channel, and every channel passes the same permission checks.
Employee assistant
Sign in with the company's Microsoft account. Ask in chat. Get a cited answer and open the source beside it.
Website chat ("Tasi")
Added to the company site with one line of code. Answers from public information only, in the company's voice.
Human support desk
Takes over a chat with a written brief and the full history. The customer stays in the same chat window. Hand-back returns to Tasi.
Admin console
Manage documents and who may read them, publish the website chat, review handoffs, and preview answers as any user.
Underneath
- 01
Document understanding
Reads digital files and scanned paper, and keeps page positions so a citation opens at the right spot.
- 02
An agent that checks itself
Every answer is verified against its sources before anyone sees it.
- 03
Connected knowledge
Links people, teams and systems across documents, for questions that span several of them.
- 04
Escalation
Knows when a person should take over, and hands back with full context.
- 05
Usage guardrails
Public channels have limits on volume and spend.
How it keeps data safe
Permissions are enforced before the assistant reads anything. It cannot reveal what it never receives.
Every change runs against a security test suite: cross-company probes, customers asking for internal facts, and instructions planted inside documents. A single leak blocks the release.
Company isolation
Each company's data is separated at the database level.
Source permissions respected
Each person retrieves only what they could open in the original source.
Security decisions in code
Rules about who may do what are enforced by the platform, not by instructions to the model.
Content is data, not commands
Documents and chat messages cannot redirect the assistant.
Personal data masked
Sensitive details are removed before they reach logs or the support desk.
Graded, not demoed
Every capability was graded on a test set of questions with known answers, asked as each kind of user, with permission traps built in. Questions were written before the features they test.
Answers arrive in seconds, already fact-checked, so no one sees a draft that later fails.
| Measure | Result |
|---|---|
| Permission leaks | 0 in every run |
| Answer correctness | 99 to 100% |
| Correct "I can't find that" answers | 97% |
| Prompt-injection attempts declined | All |
| Escalations to a person that should fire | All |
| Words read correctly from scanned pages | 96% |
| Personal data reaching logs | None |
The discipline behind it
Questions before code
Each feature starts with the test questions it must pass.
Every component earns its place
We keep a part only if it measurably improves results, and we say so when it falls short.
Security tests block releases
A single leak stops a change from shipping.
Hands-on testing
Live sessions as staff, customers and support agents shaped the final design.
Built to run lean
Running cost was a design requirement from day one.
Built with
An enterprise stack chosen for security, portability and cost control. Infrastructure is defined as code, deployments are automated, and every change is evaluated before it ships.
- Microsoft Azure
- Microsoft Entra ID
- Next.js
- TypeScript
- Python
- LangGraph
- PostgreSQL
- Chatwoot
- Langfuse
What's next
Actions, in progress
Bookings, tickets and other business-system tasks, each confirmed by the user before it runs.
More sign-in options and channels
Including Microsoft Teams.
Voice
The same assistant, permissions and handoff over the phone.
Work with Quantic Systems
We built TasiNexus end to end: cloud infrastructure as code, enterprise sign-in, document processing, retrieval, agent orchestration, evaluation and the product UI.
Run TasiNexus for your company
Your documents, your directory, your website. Permissions follow the ones you already have.
Build the same capability into your product
We bring the architecture, the security model and the test discipline shown here.
AI copilots and agents →