Product
Unify. Watch. Resolve.
Three verbs, one graph. Each one shows its work.
Unify
One sentence tells you what happened.
Drop a CSV or Excel file and CerebriNode reads a sample first: the columns and their types, the one that identifies each row, and any that look sensitive. You decide, then it imports, keeps any row that fails a check aside with the reason, and writes the graph. It tells you what it recorded and what it left out.
Links between files are proposed from matching values and confirmed by you. There is no language model in that step, so a proposed link is a fact about your data, not a guess.
- Today: CSV, Excel, Parquet, JSON, PDF text, PostgreSQL (the sync only reads)
- Next: Google Sheets, MySQL, SQL Server, MongoDB, Snowflake, BigQuery, Salesforce, HubSpot, SAP, NetSuite, Kafka, change capture planned
- Sync: re-upload a file, or connect PostgreSQL and choose how often it syncs: once, or every 1, 6, 12 or 24 hours.
suppliers.csv became 40 Suppliers, keyed by supplier_id. contact_email looked like email addresses, and you chose to leave it out.
| column | recorded as | example |
|---|---|---|
| supplier_id | key | SUP-0412 |
| name | property | Kestrel Components |
| part_id | link → Part | P-88102 |
| contact_email | left out, your choice | not kept |
Illustration, sample data
Open recalls with no vehicle linked
Sweeps hourly · delivers to #quality on Slack · a finding when the count changes
| sweep | count | change | finding |
|---|---|---|---|
| 09:00 | 4 | 0 | - |
| 10:00 | 4 | 0 | - |
| 11:00 | 6 | +2 | high |
Illustration, sample data
Watch
A question that keeps asking itself.
A watch is a question with a schedule. Watchtower runs it on the schedule you set, keeps the count, and writes a finding when the answer meets the condition you set, with the query and the counts behind it. Findings go where your team already looks: email, Slack, Teams or a webhook.
Your organization has a monthly spend cap you set. When the cap is reached, sweeps pause and the product says so. Nothing runs quietly.
- Findings carry the query’s result, severity, and how many sweeps have seen them
- Acknowledge, resolve or mute from the ledger, or ask Axon what to check
- Every sweep and every model call appears in the spend ledger with its cost
Resolve
CerebriNode proposes. You approve. It is applied.
For a missing or broken link, or the same thing recorded twice, CerebriNode proposes a fix: what will change, what it would touch, and whether it can be undone. Approve it and it is applied inside CerebriNode with a receipt: each record before and after, with your name on the approval. Your team reviews fixes instead of hunting for them.
- Today: link the missing record, mark a broken reference, join two records that are the same thing, mark a watch result as expected or move its line, hand off by webhook to systems you already run
- Next: update the row in your database, open the ticket, email the supplier, post to Slack and act on the reply planned
- One approval per proposal, unless an owner or admin turns on auto-run for a kind of fix on a screen that states the risk. Sensitive columns always ask.
highOpen recall with no vehicle linked
RC-2041 · part P-88102 · found 10:00 · 6 rows
Proposed resolution Link RC-2041 to the 6 vehicles whose VINs it lists: each VIN matches exactly one vehicle.
Dry run: 6 vehicles matched by VIN. 0 conflicts. Reversible.
10:24 approved by Priya · 10:24 6 links written, with a receipt · 11:00 sweep: count back to 4, no finding
Illustration, sample data
Ask
Axon, the agent that acts with you.
Ask Axon to create a watch, decide a proposed link, or acknowledge, resolve or mute a finding. It shows a card with the change, you confirm, and where the change can be undone, Undo is one click. It never writes without you, and everything it did is marked “through Axon” in the audit log.
Watch the recalls that don’t link to a vehicle and tell me when the count changes.
I can set that up as a watch.
Illustration, sample data
Also in the product
Knowledge graph
A WebGL field of every entity and link in your organization. Search, jump to a record, and see what it connects to.
Four roles
Owner, admin, member, viewer. Viewers can ask Axon; only the roles that can write can confirm.
Receipts and audit log
Every change to your records has a receipt: what changed, by whom, and it cannot be edited. The audit log records decisions and settings, through what.
Delivery channels
Email, Slack, Microsoft Teams and webhooks, tested from the product before a watch uses them.
MCP
One MCP address your AI tools sign in to. A person signs in, picks the organization and approves; the tool reads the graph, never changes it, and stops the moment it is disconnected. No keys.
Export
An owner or admin can download a JSON export from Settings at any time: the graph up to 10,000 records and 50,000 links, each source's settings, the members, and up to 10,000 audit log rows from the last 90 days. The rows the Data Hub stores and uploaded files are not in it.