The Company Brain

Ask why. Get a cited answer, not another dashboard.

Forty tools, resolved into one graph.

Why a dashboard can't answer "why"

A dashboard shows you the drop. It can't tell you the cause.

So you pivot. Cross-reference. Open Stripe , then Zendesk , then Amplitude , then Slack . You stitch a story by hand and call it a hypothesis. It takes an afternoon, and it's still a guess.

Vernais closes that loop. It reads all forty tools at once, resolves them into one graph, then reasons across the graph to find the cause, and hands you the receipt for every claim.

Dashboards tell you what changed.
Corvex tells you why it changed, and proves it.

01 · The Company Brain

Every tool you use, resolved into one graph you can query.

Your data doesn't live in one place. Stripe knows about charges. Zendesk knows about tickets. GitHub knows about commits. None of them know about each other.

The Company Brain does. It ingests all forty sources, cleans and deduplicates every record, then resolves the shared things, the same customer, the same version, the same service, into single nodes. What was scattered across silos becomes one connected map of your product reality.

Three node types, one meaning

signal

Signal

One clean, deduplicated record: a Jira issue, a Slack message, a Stripe charge. The atomic unit of evidence. Every Signal carries its source, its exact text, and its provenance.

entity

Entity

A resolved real-world thing that appears once, no matter how many tools mention it. Acme Corp in Stripe , acme.com in Zendesk , and Acme in a Slack thread collapse into one Entity node. This cross-tool resolution is the glue: it's what lets an investigation walk from a failed charge, to the ticket that reported it, to the commit that caused it.

topic

Topic

An ML-discovered cluster of related Signals, organized umbrella → subtopic. Onboarding & setup issues is an umbrella; Invite-code failure at onboarding is a subtopic under it. Topics are the map's index; they scope every question so the AI never has to read the whole graph.

02 · The Investigator

Ask "why" in plain language. It reasons; it doesn't search.

Type a question the way you'd ask a senior analyst: "Why are activations dropping?"

The Investigator doesn't search. It reasons. It runs an eleven-stage engine that decomposes the question, scopes it to the right corner of the graph, recalls the relevant nodes, judges each one, argues both sides, and writes a cited answer.

R0

Decompose

Breaks your question into three to five precise sub-questions.

R2

Scope

Matches the question to the right Topics and anchor Signals, so it reasons over dozens of nodes instead of hundreds of thousands.

R4

Recall

Runs Personalized PageRank from the anchors, ranking every node by relevance and keeping the bounded, on-topic set.

R6

Node-Jury

Scores each surviving node deterministically, then votes on it three times. A node only counts if it clears the relevance floor, wins a majority, and comes with a verbatim evidence span. No span, no vote.

R9

Debate

A PM voice and an Engineer voice argue the competing hypotheses against each other, surfacing counter-evidence before it commits.

R11

Synthesize

Writes the cited root-cause case file, then runs it through an entailment gate before returning it.

A full investigation takes roughly 45 to 60 seconds. It's doing real work; you see every step of it.

03 · The case file

The answer arrives as a document, not a sentence.

When the Investigator finishes, it hands you a case file. Every part of it is grounded, and every claim links back to the Signal it came from.

Root cause

The single most-supported explanation, stated plainly, and checked against its own evidence before it's shown.

Ranked hypotheses

The competing explanations it weighed, ordered by support, each with the evidence for and against. You see the runners-up, not just the winner.

Measured impact

The real numbers behind the finding: how many events, how many users, over what window, each carrying its filter recipe.

Honest confidence

High, medium, or low, and it earns the word. A causal claim is capped at "medium" unless the source text verbatim entails it. The confidence doesn't flip between runs.

Clickable citations

Every claim is a link. Tap it and land on the exact Signal, the exact source span. Nothing to take on faith.

This same case file becomes a living Initiative in one click: root cause, hypotheses, impact, and the derived metric all carry over. That's where the finding turns into work your team runs.

04 · Measured, not guessed

Every number carries its own receipt.

Most tools give you a number and ask you to trust it. Vernais gives you the number and the exact filter recipe that produced it: the query, the matched-value breakdown, a sample of the rows. No silent transforms. No quiet truncation. If you want to check the math, the math is right there.

Every count returns its recipe

The filter, the sum, and what it counted, in mono.

~917,623

Scale-verified at ~917,623 rows

Counts stay exact, and a cross-tool pair query still returns in seconds.

The North-Star metric, grounded by a closed menu

At the moment it finds the cause, the Investigator derives a single North-Star tracking metric for the problem, with the full formula and the value computed then and there.

It cannot invent that metric. The events it may choose from are a closed menu of what the field index actually discovered in your data. The numerator and denominator sum only grounded values. If a metric can't be built from real events, it isn't offered.

You lock it as a baseline, then watch it grow after you ship. The receipt travels with it.

05 · Six lanes, one chat box

One question. The right tool picks itself.

Chat is the single entry point. A router reads each message and sends it down exactly one lane, with no menus and no mode-switching.

lane 01

Deep Investigation

The full R0→R11 engine with the live stage timeline. For every "why."

lane 02

Data Lookup

A fresh number or slice from your connected data, consistent across all forty integrations.

lane 03

Web Search

Current external facts, ranked sources, distilled pages, inline citations.

lane 04

Trawl

A confirm-first conversational scrape that reverse-prompts your settings, gathers quality URLs, and streams live status. Learn how →

lane 05

Agentic

Chains graph, web, visualization, and investigation in a single turn; builds dashboards and tappable choice cards inline.

lane 06

Codebase Analyst

Semantic code search plus read-only tools, plan-first, for questions about the code itself.

06 · Grounded by design

Every fact has a receipt. Nothing is invented.

This isn't a disclaimer. It's the architecture.

Facts are stored with their source.

Every node and edge in the graph carries a source span, a Signal ID, a confidence score, and which check it passed. If it isn't grounded, it doesn't enter the graph.

Causality is derived, never baked in.

"X caused Y" is not a fact the graph stores; the data rarely says it literally. The graph holds only the grounded pieces (both Signals share Entity v2.2.1; the error spike sits TEMPORAL_NEAR the deploy). The causal story is computed fresh at query time, from the timeline, as an evidence-backed hypothesis. It is checked against that evidence before you see it.

It says "I don't know."

When there's no grounded evidence, it tells you so, plainly, instead of fabricating an answer. Off-domain questions are refused. A question that names one tool stays anchored to that tool.

The same question returns the same answer.

Seed-locked jury. Hypotheses cached to the data. Reproducible cosine scores. Ask twice, get the same route, the same confidence, the same root cause.

07 · See it all

The graph is yours to explore.

You don't have to take Corvex's word for its structure. You can walk it yourself.

Graph Explorer

An interactive, force-directed view of the entire knowledge graph. Color by entity type or integration. Click a node to highlight its neighbors. Open the inspection panel to read the full Signal text, its source payload, and everything it connects to. Dark and light.

Data browser

A full-width, filterable table over every node in the graph: Signals, Entities, Topics. Search, facet by integration and entity type, page through, and open any row's detail drawer to see the complete record and its resolved neighbors.

Data-flow walkthrough

After a Brain run, trace any real Signal's actual journey: raw payload → dedup → embed → route → extract → verify → resolve → topic, stage by stage, with the real input and output at each step. It's the proof of how a raw record became a graph node.