Home · Work

Systems Work · No. 20

Financial Crime Identity Resolution

Rajesh Ranjan Mahapatra

An autonomous multi-agent investigation that works the ambiguous residual of list screening and feeds a deterministic decision gate, with every disposition made by a human investigator.

Start reading
Systems WorkCase study

© 2026 Rajesh Ranjan Mahapatra. 6 min read

Financial Crime Identity ResolutionAll Systems Work
Home · Work

Financial Crime Identity Resolution

An autonomous multi-agent investigation that works the ambiguous residual of list screening and feeds a deterministic decision gate, with every disposition made by a human investigator.

Challenge

Screening lists and politically-exposed-person screening clears the easy alerts and leaves a stubborn residual where certainty runs out: the same common name, a partial date of birth, an ownership chain that may or may not connect a customer to a sanctioned official.

Architectural move

Five candidate architectures were evaluated against weighted criteria and a set of binary boundary tests.

Control boundary

Every claim maps to a source.

Outcome

Analysts move from clearing volume to working the genuinely hard cases, with the investigation done and cited before they open the file.

FIG.00 // SYSTEM MAPSTORE / RECORDGATE / HUMANAGENTDETERMINISTICSIGNAL / DATAIN · 01Screening signalsnames, aliases, lists,ownership data, casequeuesDET · 02Deterministicscreeningcloses the clear cases,no model involvedAGT · 03Identity inferenceagentresolves identity,traces ownership, readsadverse mediaREC · 04Cited evidence packevery claim maps to aretrieved sourceAGT · 05Independent reviewerre-derives from the packalone, never sees thereasoningGATE · 06Deterministic gatecalibrated score,independent signalsmust corroborateHUM · 07Human investigatordecides every match andevery escalationRESIDUAL ONLYEVIDENCECORROBORATEDRAIL · GWCentral model gateway on every model callprompt management · model routing · data scanning · token metering · long-retention audit · hard per-case cost ceilingSYS/FCIR · FIG.00 · SCHEMATIC · NOT TO SCALE
FIG.00 // SYSTEM MAPSIGNAL / DATADETERMINISTICAGENTGATE / HUMANSTORE / RECORDIN · 01Screening signalsnames, aliases, lists, ownership data, casequeuesDET · 02Deterministic screeningcloses the clear cases, no model involvedRESIDUAL ONLYAGT · 03Identity inference agentresolves identity, traces ownership, readsadverse mediaEVIDENCEREC · 04Cited evidence packevery claim maps to a retrieved sourceAGT · 05Independent reviewerre-derives from the pack alone, never sees thereasoningCORROBORATEDGATE · 06Deterministic gatecalibrated score, independent signals mustcorroborateHUM · 07Human investigatordecides every match and every escalationRAIL · GWCentral model gateway on every modelcallprompt management · model routing · datascanning · token metering · long-retentionaudit · hard per-case cost ceilingSYS/FCIR · FIG.00 · SCHEMATIC
FIG.00 System map. Signals pass deterministic screening first; only the residual reaches the agents, and every disposition is fixed by a deterministic gate before a person decides.

The problem

FIG.01 // THE PROBLEMREJECTED PATHHUMAN AUTHORITYDETERMINISTICSIGNAL / DATAQ-00Screening alertsSanctions lists andpolitically-exposed-person screening onevery customerQ-01Easy alerts cleared by rulesDefensible to a regulator, but arules-only control cannot investigateambiguityQ-02Stubborn residualCommon name · partial date of birth · anownership chain that may or may not reacha sanctioned officialQ-03Worked entirely by handWhere cost, risk and regulatory exposureconcentrate; the queue never shrinksSYS/FCIR · FIG.01 · SCHEMATIC · NOT TO SCALE
FIG.01 // THE PROBLEMSIGNAL / DATADETERMINISTICHUMAN AUTHORITYREJECTED PATHQ-00Screening alertsSanctions lists and politically-exposed-personscreening on every customerQ-01Easy alerts cleared by rulesDefensible to a regulator, but a rules-onlycontrol cannot investigate ambiguityQ-02Stubborn residualCommon name · partial date of birth · anownership chain that may or may not reach asanctioned officialQ-03Worked entirely by handWhere cost, risk and regulatory exposureconcentrate; the queue never shrinksSYS/FCIR · FIG.01 · SCHEMATIC
FIG.01 The problem. Rules clear the easy alerts; the residual where certainty runs out was worked by hand, and that is where cost and risk concentrate. Widths are indicative.

Screening lists and politically-exposed-person screening clears the easy alerts and leaves a stubborn residual where certainty runs out: the same common name, a partial date of birth, an ownership chain that may or may not connect a customer to a sanctioned official. At a large regulated enterprise, that residual was worked entirely by hand, and it is where cost, risk and regulatory exposure concentrate. A rules-only control is defensible to a regulator but cannot investigate ambiguity, so the manual queue never shrinks.

I designed the architecture: an autonomous multi-agent investigation feeding a deterministic decision gate. Agency lives in the investigation; the verdict stays deterministic.

Why it was hard

FIG.02 // WHY IT WAS HARDREJECTED PATHGATEAGENTSIGNAL / DATAFAIL · ALet a model decide matchesan unexplainable control in themost scrutinised corner of theenterpriseDESIGNHold both at onceinvestigative autonomy for reading,tracing and correlating; mechanicalreproducibility for decidingFAIL · BKeep everythingrules-basedthe residual stays manualforever→ EXPLAINABILITY LOSTQUEUE NEVER SHRINKS ←TERRAIN EVERY OPTION MUST SURVIVEHZ · 01Adversarial identitiesaliases, layered corporate structures,family relationshipsHZ · 02Context decayearly errors and stale retrievalscompound through a growing contextHZ · 03Model risk scrutinyreproducible dispositions, evidenceper claim, a person accountable permatchSYS/FCIR · FIG.02 · SCHEMATIC · NOT TO SCALE
FIG.02 // WHY IT WAS HARDSIGNAL / DATAAGENTGATEREJECTED PATHFAIL · ALet a model decide matchesan unexplainable control in the mostscrutinised corner of the enterprise↑ EXPLAINABILITY LOSTDESIGNHold both at onceinvestigative autonomy for reading, tracingand correlating; mechanical reproducibilityfor deciding↓ QUEUE NEVER SHRINKSFAIL · BKeep everything rules-basedthe residual stays manual foreverTERRAIN EVERY OPTION MUST SURVIVEHZ · 01Adversarial identitiesaliases, layered corporate structures, familyrelationshipsHZ · 02Context decayearly errors and stale retrievals compoundthrough a growing contextHZ · 03Model risk scrutinyreproducible dispositions, evidence perclaim, a person accountable per matchSYS/FCIR · FIG.02 · SCHEMATIC
FIG.02 Why it was hard. The two failure modes point in opposite directions; the design keeps autonomy in the investigation and reproducibility in the decision.

The failure modes point in opposite directions. Let a model decide matches and you have an unexplainable control in the most heavily scrutinised corner of the enterprise. Keep everything rules-based and the residual stays manual forever. The design had to hold both: genuine investigative autonomy where the work is reading, tracing and correlating, and mechanical reproducibility where the work is deciding.

Identity resolution itself is adversarial terrain. The entities worth screening hide behind aliases, layered corporate structures and family relationships, and the evidence is spread across registries, filings and adverse media of wildly uneven quality. Long investigations also degrade agent systems in a specific way: early errors and stale retrievals compound through a growing context, which is exactly what a reviewer later cannot untangle.

And every element had to survive scrutiny under model risk requirements: reproducible dispositions, evidence for every claim, a person accountable for every match.

Key decisions

FIG.03 // KEY DECISIONSGATEAGENTDETERMINISTICSIGNAL / DATAL1 · FIRST CRACKDeterministicclassificationIdentity matching onevery caseClear cases closed, nomodelOnly the residual movesonL2 · INVESTIGATEAgents on theresidualState rebuilt eachstageEstablished · open ·evidenceHistory in audit log,not promptL3 · CORROBORATEIndependent reviewerReads the evidence packonlyNever sees investigatorreasoningConvinced, notpersuadedL4 · DISPOSECalibrateddeterministic gateNames and aliasesGeographyProfessional contextNetwork relationshipsDemographicplausibilityRESIDUALEVIDENCE PACKEVIDENCE, NEVER MODEL OUTPUTSYS/FCIR · FIG.03 · SCHEMATIC · NOT TO SCALE
FIG.03 // KEY DECISIONSSIGNAL / DATADETERMINISTICAGENTGATEL1 · FIRST CRACKDeterministic classificationIdentity matching on every caseClear cases closed, no modelOnly the residual moves onL2 · INVESTIGATEAgents on the residualState rebuilt each stageEstablished · open · evidenceHistory in audit log, not promptL3 · CORROBORATEIndependent reviewerReads the evidence pack onlyNever sees investigator reasoningConvinced, not persuadedL4 · DISPOSECalibrated deterministic gateNames and aliasesGeographyProfessional contextNetwork relationshipsDemographic plausibilityRESIDUALEVIDENCE PACKEVIDENCE, NEVER MODEL OUTPUTSYS/FCIR · FIG.03 · SCHEMATIC
FIG.03 Key decisions. The one architecture of five that passed every boundary test: deterministic first, agents on the residual, an independent re-derivation, and a gate that scores five independent signal families.

Five candidate architectures were evaluated against weighted criteria and a set of binary boundary tests. One option, and only one, passed every boundary test and led the weighted evaluation: a staged pipeline in which deterministic classification and identity matching take first crack at every case, and only the residual they cannot close reaches the agents.

The disposition is fixed by a deterministic gate: a transparent, calibrated score over independent signal families spanning names and aliases, geography, professional context, network relationships and demographic plausibility, anchored so that several independent signals must corroborate before any recommendation stands. No model output feeds the gate directly, only evidence.

Case state is reconstructive rather than cumulative. Each stage rebuilds a minimal sufficient state from the governed case object: what is established, what is open, what evidence supports each claim, nothing else. Full history lives in the audit log, never in the prompt.

Before a true-match recommendation reaches the gate, an independent reviewer agent re-derives it from the evidence pack alone, without sight of the investigator agent's reasoning. It cannot be persuaded, only convinced by the same facts, which turns agreement into corroboration rather than compliance.

Architecture

FIG.04 // ARCHITECTUREREJECTED PATHSTORE / RECORDAGENTDETERMINISTICSIGNAL / DATAAGTAgent drafts a claimidentity, ownership oradverse-media findingCHK · 01Hallucination checkunsupported claims arestrippedCHK · 02Citation checkthe claim must mapmechanically to aretrieved sourceRECCited evidence packonly checked claims arewrittenREJECTLOOPBad citationforces regeneration of the claimRAIL · GWEvery model call crosses the central gatewayprompt management · model routing · data scanning · token metering · long-retention audit · hard per-case cost ceilingSYS/FCIR · FIG.04 · SCHEMATIC · NOT TO SCALE
FIG.04 // ARCHITECTUREDETERMINISTICAGENTSTORE / RECORDREJECTED PATHAGTAgent drafts a claimidentity, ownership or adverse-media findingCHK · 01Hallucination checkunsupported claims are strippedCHK · 02Citation checkthe claim must map mechanically to a retrievedsourceRECCited evidence packonly checked claims are writtenREJECTLOOPBad citationforces regeneration of the claimRAIL · GWEvery model call crosses the centralgatewayprompt management · model routing · datascanning · token metering · long-retentionaudit · hard per-case cost ceilingSYS/FCIR · FIG.04 · SCHEMATIC
FIG.04 Architecture. The system cannot show a material claim unless it can say, mechanically, where the claim came from; the gateway carries cost and audit as infrastructure.

Signals arrive from screening systems and case queues. Deterministic screening closes the clear cases. The residual flows to the identity inference agent, which resolves identity ambiguity, traces beneficial ownership through corporate and family structures, reads adverse media, and assembles a cited evidence pack in which every claim maps to a retrieved source. The independent reviewer re-derives the recommendation, the deterministic gate scores it, and a human investigator decides every match and every escalation.

All model calls cross a central gateway carrying prompt management, model routing, data scanning, token metering, long-retention audit and a hard per-case cost ceiling. A hallucination check strips unsupported claims, and a bad citation forces regeneration: the system cannot show a material claim unless it can say, mechanically, where the claim came from.

Trade-offs

FIG.05 // TRADE-OFFSREJECTED PATHGATEDETERMINISTIC▼ PAYSDESIGN CHOICE▲ BUYSCaps how clever the systemmay be: fluent matchreasoning never becomes averdictT-01Deterministic gateEvery dispositionreproducible from the caseobject and the scoreAssembly work at every stageT-02Reconstruction overaccumulationRoughly linear token cost, nocontamination, any stagereplayableRoughly doubles inference onhigh-consequencerecommendationsT-03Independent reviewerAgreement becomescorroboration: cheapinsuranceTHE CAP IS THE FEATURESYS/FCIR · FIG.05 · SCHEMATIC · NOT TO SCALE
FIG.05 // TRADE-OFFSDETERMINISTICGATEREJECTED PATHT-01Deterministic gate▼ PAYSCaps how cleverthe system maybe: fluent matchreasoning neverbecomes averdict▲ BUYSEverydispositionreproduciblefrom the caseobject and thescoreT-02Reconstruction over accumulation▼ PAYSAssembly work atevery stage▲ BUYSRoughly lineartoken cost, nocontamination,any stagereplayableT-03Independent reviewer▼ PAYSRoughly doublesinference onhigh-consequencerecommendations▲ BUYSAgreementbecomescorroboration:cheap insuranceTHE CAP IS THE FEATURESYS/FCIR · FIG.05 · SCHEMATIC
FIG.05 Trade-offs. Each choice pays a visible cost for a property a model risk reviewer asks for.

The deterministic gate caps how clever the system is allowed to be, and that cap is the feature: fluent reasoning about a possible match is investigation material, never a verdict. Reconstruction over accumulation costs assembly work at every stage, but it keeps token cost roughly linear across a long investigation, stops one bad retrieval contaminating every later step, and makes any stage replayable in isolation, which is the property a model risk reviewer asks for and rarely gets. The independent reviewer roughly doubles inference on high-consequence recommendations; on a restricted parties-adjacent decision, that is cheap insurance.

Governance

FIG.06 // GOVERNANCESTORE / RECORDGATE / HUMANDETERMINISTICCONTROLMECHANISMSTATEEvery claim maps to a sourceHallucination check strips; a bad citation forcesregenerationBLOCKINGEvery disposition is reproducibleRebuilt from the case object and the calibrated scoreREPLAYABLEEvery positive and escalation goesto a personA human investigator decides; agents only recommendHUMAN-OWNEDThe decision log is auditable endto endLong-retention audit captured at the gatewayAUDITEDCost ceiling and audit trailEnforced as gateway infrastructure; no build can opt outENFORCEDSYS/FCIR · FIG.06 · SCHEMATIC · NOT TO SCALE
FIG.06 // GOVERNANCEDETERMINISTICGATE / HUMANSTORE / RECORDBLOCKINGEvery claim maps to a sourceHallucination check strips; a bad citationforces regenerationREPLAYABLEEvery disposition is reproducibleRebuilt from the case object and thecalibrated scoreHUMAN-OWNEDEvery positive and escalation goes toa personA human investigator decides; agents onlyrecommendAUDITEDThe decision log is auditable end toendLong-retention audit captured at the gatewayENFORCEDCost ceiling and audit trailEnforced as gateway infrastructure; no buildcan opt outSYS/FCIR · FIG.06 · SCHEMATIC
FIG.06 Governance. Explainability as a data structure, not as prose written after the fact.

Every claim maps to a source. Every disposition is reproducible from the case object and the calibrated score. Every positive and every escalation routes to a person, and the decision log is auditable end to end. The gateway enforces the cost ceiling and the audit trail as infrastructure rather than convention, so no individual build can quietly opt out. This is the shape financial regulators keep asking for: explainability as a data structure, not as prose written after the fact.

Outcome

FIG.07 // OUTCOMEDETERMINISTICSIGNAL / DATAANALYST TIMEclearing volumeHard casesWHERE THE ANALYSTS NOW WORKCASE FILEinvestigate on openCited firstINVESTIGATION DONE BEFORE THE FILE OPENSSAME EVIDENCE-FIRST ARCHITECTURE, EXTENDING ACROSS THE FINANCIAL-CRIME ESTATESuspicious-activity report draftingRestricted-parties decision supportFraud-dispute agent assistanceAGENTS RECOMMEND. HUMANS DECIDE. EVERYWHERE.SYS/FCIR · FIG.07 · SCHEMATIC · NOT TO SCALE
FIG.07 // OUTCOMESIGNAL / DATADETERMINISTICANALYST TIMEclearing volumeHard casesWHERE THE ANALYSTS NOW WORKCASE FILEinvestigate on openCited firstINVESTIGATION DONE BEFORE THE FILE OPENSSAME EVIDENCE-FIRST ARCHITECTURE,EXTENDING ACROSS THE FINANCIAL-CRIMEESTATESuspicious-activity report draftingRestricted-parties decision supportFraud-dispute agent assistanceAGENTS RECOMMEND. HUMANS DECIDE.EVERYWHERE.SYS/FCIR · FIG.07 · SCHEMATIC
FIG.07 Outcome. Analysts move from clearing volume to the genuinely hard cases, and the pattern is spreading across the estate.

Analysts move from clearing volume to working the genuinely hard cases, with the investigation done and cited before they open the file. The pattern is extending across the financial-crime estate on the same evidence-first architecture: drafting of suspicious-activity reports, restricted parties decision support, and agent assistance for fraud disputes. Agents recommend and humans decide, everywhere.

What I learned

FIG.08 // WHAT I LEARNEDREJECTED PATHGATEAGENTAX-01Separate agency fromauthorityThe governance fight only existswhen one component tries to holdboth.AX-02Independent derivation beatsagreementA reviewer that has read theinvestigator's reasoning can be talkedinto things.AX-03The residual is the productAnyone can automate the easy matches;value and risk live where certaintyruns out.SYS/FCIR · FIG.08 · SCHEMATIC · NOT TO SCALE
FIG.08 // WHAT I LEARNEDAGENTGATEREJECTED PATHAX-01Separate agency from authorityThe governance fight only exists when onecomponent tries to hold both.AX-02Independent derivation beats agreementA reviewer that has read the investigator'sreasoning can be talked into things.AX-03The residual is the productAnyone can automate the easy matches; valueand risk live where certainty runs out.SYS/FCIR · FIG.08 · SCHEMATIC
FIG.08 What I learned. Three lessons from the engagement.

Separate agency from authority and most of the governance argument dissolves; the fight only exists when one component tries to hold both. Independent derivation is the strongest form of agreement: a reviewer that has read the investigator's reasoning is a reviewer that can be talked into things. And in screening work, the residual is the product. Anyone can automate the easy matches; the value and the risk both live in the cases where certainty runs out.