Platform
Agents investigate through fixed agentic workflows. A deterministic engine makes the call: every rule versioned, every score reproducible. Your analysts stay in command.
The products · eight, one platform
Each one solves a specific compliance job. All eight run on the same architecture (agents gather the evidence, a deterministic engine makes the call, your analysts command) and draw on the same shared capabilities.
Wallet risk arrives through the blockchain analytics you already trust (Chainalysis, TRM, Elliptic) and enters the same versioned engine as every fiat alert. We don't read the chain. We decide the case.
See the tools that power themA clean applicant clears every deterministic check and is approved in under ninety seconds: screening included, no queue. The next application declares 'student' as occupation against an opening deposit that doesn't fit the profile: the occupation-consistency check refers it, and the analyst opens a file where the question is already framed and the evidence already attached.
REQUEST THE SCREENING WALKTHROUGHSeven declared claims behind one private-banking relationship. Six are matched to registrar, banking and title records before lunch. The seventh, a private loan repayment, has no record to match: the case clock pauses, a bank-approved document request goes out, and the memorandum that used to take two weeks ships the same week, honest about which claim arrived last.
REQUEST THE MEMORANDUM WALKTHROUGHOvernight, the batch reads 240,000 settled transactions. By sunrise there are thirty-eight alerts: ranked, not piled, each carrying the rule that fired with its thresholds quoted and a narrative already written, every claim cited. The analyst's morning starts at alert #1, not alert #400.
REQUEST THE MORNING-QUEUE WALKTHROUGHA salary account with one employer deposit a month, three years running, suddenly receives transfers from fourteen unrelated senders in a single week. Against the general population, small sums, no alarm. Against this customer's own baseline, the departure is unmistakable, and the alert says so in exactly those terms, baseline attached.
SEE THE BASELINESOne account receives a large credit and splits it across nine new payees within two hours, then does it again the next week. Per transaction, each transfer looks routine. Scored at the customer, the pass-through share and payee velocity tell a different story, and the precedent search finds the same shape confirmed in a past case. The file opens with the pattern already drawn.
REQUEST THE MULE-RING WALKTHROUGHA six-month bank statement: forty scanned pages and a handwritten annexure. The agent reads it into a structured ledger, every line carrying its page and region, so when the narrative cites three specific deposits, each citation opens the exact rows it came from. What can't be read cleanly is flagged for a human, never guessed.
REQUEST THE DOCUMENT WALKTHROUGHA pattern keeps slipping past the rulebook: rapid pass-through just under the thresholds, always via newly added payees. The system drafts the missing rule in the same readable grammar as the other nine, backtest evidence attached, and routes it to one place: the MLRO's signature. Nothing activates until a human signs, and once signed, it's versioned like everything else.
SEE THE KNOWLEDGE HUBA withdrawal is instructed to an external wallet, and the address is an exact match to a listed entry. No score, no weighing: the mandate lane goes past the model and the score entirely, straight to a named human, above the queue — with the match evidence and receipts already filed. The same customer's fiat history and wallet evidence sit in one case file when they get there. Mandates are law, not machine learning.
REQUEST THE TWO-RAIL WALKTHROUGHArchitecture · one governed flow
Two intake rails, banking and virtual assets, feed one governed architecture. Agents investigate, a deterministic engine decides, your analysts command. Two limits are drawn into the picture below: when the rules are already confident the language model is skipped, and a sanctions mandate goes to a named human instead of into a score.
The gate · deterministic rules run first
Rules confident → language model skipped entirely
Rules unsure → language model gathers & corroborates, then hands evidence back
An exact watchlist match is never turned into a score. It skips the model and the score — never the evidence: deterministic gathering still files the match record and its receipts — and is held for mandatory review by a named person, above the ranked queue.
no model · no score · evidence still filed
Investigate. Decide. Command. In that order, every time.
Seven agentic specialists gather the evidence: names, graphs, registries, documents, media, precedent. Language models do the reading (resolving name variants, extracting scanned documents, sweeping adverse media) and every finding is tied to the record behind it. What a model can't corroborate is flagged, never asserted.
A deterministic engine fuses every signal under versioned weights. Same inputs, same score, every time. When the rules are confident, the model is never called at all.
Every disposition takes a reason code and a countersignature the schema physically requires. Confirmed suspicion produces a verified STR narrative and a goAML-aligned pack for the MLRO to file. The score is a recommendation; the decision is your analyst's.
Bring up to 5,000 historical alerts. Two weeks later you're reading your own numbers, not our brochure.