Author
Compose SwitchRules, decision tables, and Compute nodes in the browser. CEL expressions for predicates – sandboxed, fast, familiar to anyone who has touched Kubernetes.
Cerebra
Decisions, deployed in milliseconds
A modern BRMS your risk team and developers both like. Author business rules visually. Version like code. Serve via APIs with single-digit-millisecond latency.
Book a demo
The Problem
Business rules are the most important code at your company, and no one treats them like code.
01
Legacy BRMS are slow to change.
A risk team writes a rule. Engineering writes a JIRA. Six weeks later, an appliance ships. By then the policy is wrong, the consultant is paid, and the deal is dead.
02
Spreadsheets and code drift apart.
The policy doc says one thing. The Python service does another. The Excel sheet underwriters actually use does a third. No one knows which one decided.
03
Auditors want a story you can’t tell.
“Why did you decline this customer on March 4th?” The honest answer involves a deploy log, a deleted branch, and a developer who left in May.
How It Works
Three primitives. One request path.
Publish
v3.149f4ea201…LIVE
v3.137c12bb83…ROLLBACK
v3.12a90f1e44…
v3.11b521e93c…
Cerebra compiles the graph to an in-memory program, hashes it with SHA-256, and writes the result to an append-only registry. A reviewer signs off. The pointer flips.
Evaluate
P500.9ms
P952.4ms
P994.2ms
Your service POSTs JSON. Cerebra resolves the live registry, runs the program lock-free, returns a decision and the exact node path that produced it. Every call is recorded.
The Landscape
What you replace, and what you stop building.
Cerebra
DIY service
Authoring
Visual graph + CEL
Eclipse-based BOM tooling
Excel + proprietary DSL
Whatever you can ship in a quarter
Deployment
Managed SaaS, multi-region
WAS appliance
On-prem appliance
Your monorepo
Versioning
SHA-256, atomic swap
Project import/export
Manual snapshots
git tag, fingers crossed
Audit trail
Per-call, append-only
Add-on module
Add-on module
You’ll get to it
P99 latency
~4 ms
up to 250 ms
up to 80 ms
Depends on the intern
Runtime
Rust, no JVM
JVM, proprietary
JVM, proprietary
JVM, proprietary
Pricing posture
SaaS · usage-based · $499/mo
$1M+ & consultants
$500K+ license & SI
Headcount
First rule shipped
An afternoon
Two quarters
One quarter
“Q3, maybe Q4”
Built for Regulated Workloads
Every decision has a paper trail.
Yours can have one too.
Lending, KYC, fraud, claims, eligibility. The decisions auditors care about most are the ones engineers have the least time to instrument.
- Four-eyes review
- Every publish requires a second signer. The signer is recorded with the version.
- Replay against draft
- Pipe the last 24h of production traffic at a draft graph. See what would have changed before it changes.
- Atomic rollback
- Any prior version is one click away. The swap is lock-free; in-flight requests finish on the old version.
- Append-only log
- Input, output, fired node, exact program hash. Sign it, ship it to your warehouse, hand it to your regulator.
Architecture You Can Trust
A small, boring footprint that earns its keep in the request path.
Single static binary. No JVM, no proprietary runtime, no agent.
Battle-tested, replicated, point-in-time recoverable. The boring choice, on purpose.
One POST /api/evaluate. JSON in, decision out. Drop in.
Multi-region by default. Dedicated single-tenant regions on Enterprise. No infra for you to run.
Map your existing groups to author / reviewer / publisher.
Programs are in-memory; evaluations are lock-free atomic reads.
Where It Ships
Same primitives. Every regulated decision in the request path.
A SwitchRule is a SwitchRule whether it routes a loan or a claim. Pick an industry — see the graphs Cerebra is built to run.
Underwrite a book, not a borrower.
From application through origination to portfolio review. Risk owns the policy, engineering owns the SDK, audit owns the trail. No one owns a spreadsheet.
Throughput target10M+ / day
Latency target< 3 ms
Eligibility
Loan origination
Bureau pull → tier → instant approve / manual review.
Pricing & risk
Rate-tier pricing
APR bands by FICO, DTI, and tenure.
Fraud & AML
Application fraud
Device, velocity, and ID-document signals.
KYC / identity
Onboarding KYC
Sanctions, PEP, address-history thresholds.
Limits & auth
Credit-line set
Initial line and step-up schedule.
Workflow routing
Manual review
Queue assignment by complexity and SLA.
Flagship example we demoSupported · build it in an afternoonNot applicable in this industry
In the Wild
Three shapes of policy. The same three primitives.
Pull bureau and income, tier the applicant, route to instant-approve or manual review. 47 rules. One graph. Underwriters edit it themselves.
Decisions/day1.2M
P992.8 ms
AuthorRisk team
Velocity, device, geo, merchant category, and a model score. Six branches, two outcomes — pass or hold for review. Every threshold is a version.
Decisions/day38M
P991.4 ms
AuthorFraud ops
Geography, age, occupation, coverage limits. The graph reads like the underwriting manual — because the underwriting team wrote it.
Decisions/day96K
P994.6 ms
AuthorUnderwriting
Developer Experience
It’s just an API.
No SDK to vendor. No DSL to learn. No agent to install. POST a payload, read a decision. The reverse-loop is the same — push a graph, get back a version hash.
- SDK
- Optional. Generated from the OpenAPI spec.
- Tracing
- OpenTelemetry on every evaluate call.
- Local
docker run cerebra/devon your laptop.- CI
- Diff a draft against prod. Fail the build if behavior changes.
$ curl -X POST https://api.cerebra.io/v1/evaluate \ -H "Authorization: Bearer $CEREBRA_KEY" \ -H "Content-Type: application/json" \ -d '{ "graph": "policy.loan.origination", "input": { "applicant": { "income": 92000, "dob": "1989-04-12" }, "bureau": { "fico": 742, "delinq_24m": 0 } } }'{ "decision": "approve", "tier": "prime", "fired_node": "tier.route#3", "node_path": ["applicant", "tier.route", "approve"], "version": "3.14", "program_sha": "9f4ea201bc54…", "latency_ms": 3.4}Pricing
Priced like infrastructure, not like an appliance.
No per-seat tax on your risk team. No five-figure procurement to ship one rule. Pay for evaluations and graphs — the things that actually scale with your business.
Basic
For teams shipping their first decisions in production.
$499/ month
Start free trial- Up to 3 decision graphs
- 500K evaluations / month
- 30-day audit retention
- Visual graph authoring
- CEL expressions, replay sandbox
- Slack-channel support
Premium
For production at scale, with the controls auditors ask for.
$2,400/ month
Start free trial- Up to 25 decision graphs
- 25M evaluations / month
- 1-year audit retention + warehouse export
- SSO (OIDC + SAML)
- Four-eyes review, atomic rollback
- Production replay against draft graphs
- 99.95% SLA, priority support
Enterprise
For regulated workloads, dedicated regions, and named support.
Customannual
Talk to sales- Unlimited graphs & evaluations
- Indefinite, signed audit retention
- Dedicated single-tenant region
- SCIM, custom roles, IP allow-listing
- Custom SLA up to 99.99%
- Named SE + quarterly security review
- Procurement, MSA, DPA, your paper
All plans · SOC 2 Type II · OpenTelemetry on every call · Region: us-east, eu-west, ap-south
FAQs
Frequently Asked Questions
What is Cerebra?
Cerebra is a modern business rules management system (BRMS). Author decision logic as a visual graph, version it like code with SHA-256 hashes, and serve it over an API with single-digit-millisecond latency.
Do I have to replace my existing stack?
No. Cerebra is an orchestration layer that sits alongside your services. Your application POSTs JSON to a single /evaluate endpoint and gets a decision back — no core replacement, no appliance to run.
How is a rule change deployed?
You author in the browser, a second reviewer signs off (four-eyes review), and the live pointer flips atomically. The swap is lock-free and any prior version is one click away.
Is every decision auditable?
Yes. Every evaluation is written to an append-only log with the input, output, fired node, and exact program hash. Sign it and ship it to your warehouse or hand it to your regulator.
How fast is it?
Programs are compiled to an in-memory representation and evaluated lock-free. Typical latency is ~1ms at P50 and under ~4ms at P99.
What does it cost?
Usage-based, from $499/month. You pay for evaluations and graphs — not per-seat licenses or six-figure procurement. Enterprise plans add dedicated regions and named support.