Use Cases & Benchmarking

Built for the calls you cannot get wrong.

From exchanges to custodians to treasuries, Web3Firewall and OraBrain handle the decisions that move money and carry risk. Here is where they are used, and how we measure them against generic models and legacy tools.

Pre-execution by designYour data stays in your perimeterEvidence-backed decisions
Who It Is For

One platform, many fronts.

Wherever digital-asset value moves, Prax Labs evaluates the decision before it happens, in the context of who you are and what you are responsible for.

Exchanges & trading platforms

Screen every transaction before it is broadcast, intercept scams and fraud in real time, and keep AML and sanctions checks always on.

Custodians & wallet infrastructure

Assess signing-phase risk on the most sensitive holdings with analysis that never leaves your environment.

Treasuries & institutional holders

Track counterparty exposure and screen inbound and outbound flows against your own risk policy.

DeFi protocols & on-chain apps

Catch exploit patterns, bridge and mixer exposure, and behavioral anomalies before a contract is drained.

Compliance & risk teams

Turn raw alerts into auditable, evidence-backed decisions with Travel Rule support and SAR-ready reasoning.

Fintech & banking infrastructure

Bring on-chain intelligence into the regulated stack as digital assets reach traditional finance.

How It Works

Every decision, before it happens.

Prax Labs is built around pre-execution. The same loop runs whether the trigger is a transaction, an alert, or an autonomous agent action.

01 / Detect

Detect

A signal appears: a transaction at signing, an on-chain event, or a behavioral anomaly.

02 / Reason

Reason

OraBrain models intent in your context, on-chain and traditional data together, inside your perimeter.

03 / Decide

Decide

It returns an evidence-backed verdict, stating what is known and naming what is missing.

04 / Act

Act

Web3Firewall enforces: allow, hold, quarantine, or escalate, before anything is broadcast.

Benchmarking

How we measure OraBrain.

We do not test trivia. We evaluate OraBrain against generic cloud LLMs and legacy rule-based tools on the scenarios institutions actually face, and we score the dimensions that decide real outcomes.

Scenario 01

AML / SAR determination

Given partial transaction context, reach a defensible filing decision instead of hedging.

Scenario 02

Indexer health alert

Triage an infrastructure anomaly and separate the real signal from the noise.

Scenario 03

Wallet fraud screening

Catch fraud patterns on a live wallet without drowning analysts in false positives.

What we score
Generic LLM
Legacy tool
OraBrain
Commits to a defensible decision
Partial
Yes
Yes
Names the data it is missing
No
No
Yes
On-chain domain depth
Partial
Partial
Yes
Resists confident hallucination
No
Yes
Yes
Adapts to your policy and context
No
Partial
Yes
Data never leaves your perimeter
No
Yes
Yes

We publish methodology and we show real results. The evaluation below is one example, graded by frontier models acting as judge. Full results against your own scenarios are shared during evaluation, so you can verify them rather than take them on faith.

Independent Evaluation

Judged against frontier models.

We posed a tough AML and SAR determination to OraBrain and the leading frontier models, then asked frontier models to grade the answers. Across different judges, OraBrain came out on top for this task.

Scored by a frontier model acting as judge

Model
Diagnostic Rigor
Uncertainty Handling
AML Reasoning
Overall
OraBrain
9
9
8
8.5 / 10
Claude
8
7
9
8.0 / 10
Gemini
7
4
8
6.5 / 10
OraBrain behaves like an investigator gathering evidence rather than a rules engine reaching a conclusion.Frontier model used as judge

A second judge, comparing the same class of task across the dimensions that matter

Dimension
OpenAI
Claude
OraBrain
Risk identification
Full
Full
Full
SAR decisiveness
Defers
Presumptive file
Conditional and decisive
Operational next steps
None
Light
Concrete: freeze, scan, OOB verify
Names missing data
No
Implied
Explicit
On-chain domain depth
None
None
Mixers, bridges, cross-chain, exploit sigs
Legal standard cited
No
BSA suspicion prong
Implied, not named

We show the misses too.

On a separate infrastructure-triage question, OraBrain's fast Instant tier diagnosed the issue correctly and chose the most economical first move, reading the error logs, but it under-committed on the final operational decision, which the Deep tier is built to make. Benchmarking is only useful if it is honest about both.

Talk to us

AI agents will transact financially. Static security will fail. Prax Labs is ready.