The platform
One graph. Many signals. Human judgment on top.
FraudBrain is an intelligence platform in the literal sense: it fuses records into evidence, scores what it sees, and puts a ranked queue in front of the people with the authority to act. This page describes how it thinks, at the altitude a CISO or an SIU head expects. It does not describe the mechanics.
The evidence graph
Everything becomes a node or an edge.
Applications, quotes, payments, commission ledgers and post-sale carrier evidence are fused into one graph of producers, applicants, instruments and origins. Fraud hides in connections. The graph is where connections live.
How it thinks
Records in. A record out. Six stages, one of them human.
Every producer travels this path continuously. Nothing on it is a black box to the people who use it, and nothing on it acts on a person by itself.
A live cadence, not a review season
The whole book is re-scored continuously, many times a day, and post-sale carrier evidence lands as it arrives. A pattern that forms this morning is in the queue today, not in next quarter's audit.
Append-only, sourced, timed
Every signal keeps its source and its time. Evidence is added, never overwritten, so a reviewer can always see what was known, when, and from where. The ledger is the case file.
A person, and a logged decision record
The queue ends at a human. Review, notes and the outcome are recorded with who, when and why. The record belongs to your organization and travels with the file.
Never regress on a confirmed case
Every scoring change is measured against confirmed cases before it ships. If it would lose a case the engine already catches, it does not ship. The gate blocks the release, not a person.
Identity, origin, instrument, behaviour, outcome
Signals are read across five facets: who the applicant is, where the application came from, what paid for it, how the producer behaves over time, and what happened to the policy after issue. One facet is a hint. Five, in agreement, are a case.
Carriers and BGAs, each in their own room
Every organization sees exactly its own book. The boundary is enforced server-side on every request and proven by isolation tests that block a deploy when they fail.
Plain-language questions, with the scope disclosed
Ask about a producer, a downline, or who to look at first this week. Every answer is scoped to your organization's own book, and says so, every time. It answers questions. It does not act on anyone.
Scoring
Continuously scored. Carefully shipped.
Machine-learned and engineered
Learned models score the patterns that repeat. Engineered signals encode what investigators already know. They work together, and they are allowed to disagree.
Continuous, not quarterly
The whole book is rescored around the clock. New evidence lands in the next walk of the graph, not in the next audit cycle.
The regression gate
Scoring changes are evaluated against a held-out set of confirmed cases before release. If precision on confirmed cases drops, the change does not ship.
Live-fed, aggregate-only
The production metrics on this site are aggregate counts, refreshed on a schedule and published only when the refresh fully succeeds. Nobody types a statistic here.
Humans in the loop
The system surfaces. People decide.
FraudBrain's specialists maintain the models and review what the system surfaces for quality. Decisions about producers belong to your organization: your investigators, your procedures, your record.
Advisory by design
A score determines no action. The system never clears, declines or terminates anyone. It builds the case file and ranks it; a person makes every call.
Every review teaches
Confirmed human verdicts become training signal, so the engine sharpens against threat patterns that will not sit still.
Scoped at the server
Every organization sees exactly its own book. Boundaries are enforced server-side on every request and proven by deploy-blocking isolation tests.
Beyond insurance
The engine travels.
Entities, origins, instruments, relationships. The graph does not care which industry the ledger came from. Life insurance distribution is deployment one, chosen because fraud hides best where the money moves fastest. Everything else is a data feed away.
What we will not publish
The mechanics stay off the public internet. That is a policy, not an oversight.
You will not find on this site the thresholds, the weights, the time windows, the names of individual signals, the model features, or the sources behind any signal. Every mechanic we publish is a mechanic a ring can route around, and the rings read vendor websites too. What we do publish is the shape of the system, the posture, and the guarantees: scores advise, people decide, nothing regresses on a confirmed case, every organization sees only its own book.
The full architecture, mechanics included, is walked through under agreement during the review. Bring your data people.
Ask us the hard questions.
Architecture reviews are how every deployment starts. Bring your data people.
Request an architecture reviewA human reads every request. Prefer email? hello@fraudbrain.ai