The qbrin platform
Answers you can cite. AI you can limit.
qbrin is the trust layer for people and AI agents. The company brain turns what your teams write into cited answers. Security for the AI you adopt checks each action before it runs and, in early access, limits what each AI can see.
- Cited answers
- Permission-aware
- Go, hold or ask a person
Half one · The company brain
One engine. Every answer traceable.
qbrin reads what your company already writes, types it into a knowledge graph, and compresses it once into a map that fits inside a model’s context.
- 01Your sources
Docs, PDFs, chats, sheets and email, wherever your company already writes.
- 02Typed graph
Entities extracted and typed; hubs chosen by type-weighted centrality.
- 03QWSEC compression
Compressed once into a fixed-token map that fits in a model’s context.
- 04Recovery plans
Each hub keeps a validated plan to rebuild its evidence when a question needs the detail.
- 05Verdicts
Verified with citations, rejected, or “need more evidence”, permission-checked on the web and through the API.
01 · Connect
Every scattered source, one typed graph.
qbrin reads the docs, PDFs, chats, sheets and email your teams already produce, and types what it finds: people, policies, customers, projects, rules.
- Entities extracted and typed
- Hubs picked by type-weighted centrality, so a mailing list is not mistaken for a decision
- Synced on a schedule, not on every question
02 · Compress
Compress once. Answer many times.
QWSEC (query-weighted structural-entropy compression) turns the graph into a fixed-token knowledge map that sits inside the model’s context. Questions are answered from that compact map first, with deeper retrieval only when a block needs it.
- A fixed-token map that fits in the context window
- Each hub keeps a recovery plan for the detail behind it
- Recompressed when your graph drifts, not on every question
Knowledge map · fixed token budget
03 · Cite
Every answer shows its work.
Every claim is checked against its sources before anyone sees it: Verified with citations, Rejected when the sources contradict it, or Need more evidence, with what is missing.
- Inline citations to the page, message or row
- Permission-checked against the person, or agent, asking
- A verdict instead of a confident guess
What’s our refund window for annual enterprise plans?
qbrin
Annual enterprise plans are refundable pro-rata within 60 days of the start date 1. After 60 days, refunds need CFO sign-off 2.
- 1Refund policy v3.pdfFinance · page 4 · updated 12 Aug
- 2#finance-opsSlack thread · 3 Sep
Half two · Security for the AI you adopt
Give your AI the task, not your company.
Every AI tool asks for access to everything. qbrin sits between your data and your AI, so you can see who can act, limit what each one reaches, and check what it does.
Know
See every identity that can act.
One inventory of people, AI agents, API keys, connections and unmanaged tools. Each is scored for exposure, and the fixes are ranked biggest first.
See KnowEarly accessScope
Give each AI only what its job needs.
Limit an agent, or an outside tool’s API key, to the sources its job needs. For an agent, preview what the limit would have cut from its last 30 days.
See ScopeLiveCheck
Go, hold or ask a person before it acts.
An AI proposes an action. qbrin checks it against current evidence, permissions and policy, then answers go, hold or ask a person. Every decision leaves a receipt.
See CheckCheck is live today. Know and Scope are in early access. Scope narrows by source, and it cannot narrow a tool that holds its own login to your systems: qbrin shows that as exposure instead.
Platform tour
Follow one action from registration to receipt.
Four steps show what happens when an agent tries to refund a customer. Pick a step, or let it run. Everything in the panel is illustrative.
- Credential
scoped key ••••8f2a- Tools
billing.readbilling.refundcrm.read- Trace
- Every run is recorded
The decision engine
Four checks, one verdict.
Before a tool runs, the checkpoint asks four questions. An action gets a grant only when all four are met. Anything short of that is held, or sent to a person.
- EvidenceCurrent, and tied to this exact claim
- AuthorityWho is asking
- PolicyThe tool boundary for this agent
- Injection screenInstructions hidden in the data
refund_payment₹4,200 · order #8841- EvidenceOrder and invoice match the claimMet
- AuthorityRequester verifiedMet
- PolicyInside the refund policyMet
- Injection screenNo instructions found in the dataMet
ALLOWA signed grant is issued for this action.Labs
Measured, not claimed.
Each number comes from a committed benchmark, with its sample size and its limits.
0
Made-up answers on 120 trap questions
Four corpora, with invented projects, false dates and wrong values. qbrin declined, or corrected the false premise with the cited fact. Measured 2026-07, groundedness-audited.
86%
Citations that back the answer
With the double-checking layer on, a cited source backs the answer 86% of the time. Raw keyword search manages 6%. Measured on the ERB set of real company questions.
4/4
Unsafe actions stopped
A real LangGraph agent made five merge calls. Four had missing, stale, wrong-repository or cross-tenant proof. All four ran without qbrin and none did with it, and the valid call still ran.
See qbrin on your own stack.
Book a demo to see cited answers from your own documents, and what the checkpoint does to a real agent action.