Human-AI workspace to reduce financial crime
TRACE Nexus is a working investigative workspace that fuses evidence, resolves entities, detects typologies, and drafts intelligence notes, freeing investigators to spend their time on judgement rather than paperwork. Success is measured not by the volume of detections produced, but by the number of criminal operations no longer able to run.
Detection alone does not change anything. TRACE Nexus is built to measure impact at three levels, each building on the last, because a lead confirmed in isolation rarely changes the conditions that produced it.
Level 1
Case-level disruption
An account frozen, funds recovered, a case prosecuted: the immediate, measurable result of a single investigation.
Level 2
Network-level disruption
Shared infrastructure or entities surfaced across multiple cases, so investigators dismantle a network rather than one actor at a time.
Level 3
Environmental disruption
The regulatory gaps and payment-rail vulnerabilities that let a network operate in the first place, surfaced for policymakers and regulators to close.
Borderless cybercrime succeeds because of an asymmetry in cognition. Criminal networks share infrastructure, intelligence, and payment methods across jurisdictions in real time. Law enforcement agencies operate in silos, processing fragmented evidence with tools designed for a pre-internet world.
The gap is not technological. It is cognitive. Investigators drown in data while starving for intelligence.
80% / 20%
Investigators spend roughly 80% of their time collecting, correlating, and formatting information, and only 20% doing what humans do best: interpreting context, exercising judgement, and making decisions that hold up in court.
The scale below comes from independent research, not from this project.
$800B-$2T
Estimated laundered globally every year, roughly 2 to 5% of world GDP.
UNODC estimate, via University of Maryland School of Public Policy95%+
Is not an unusual AML alert false-positive rate at financial institutions, all of it still requiring manual review.
LexisNexis Risk Solutions$20.9B
Lost to internet crime in the United States in 2025 alone, a record year, up 26% from 2024.
FBI Internet Crime Complaint Center, 2025 Annual Report$154B
Received by illicit cryptocurrency addresses in 2025, up 162% year over year.
Chainalysis, 2026 Crypto Crime ReportA symbiotic partnership where AI handles the computationally intensive work of evidence fusion, pattern detection, and cross-jurisdictional correlation, while humans handle the interpretive work: contextual understanding, ethical judgement, and operational decision-making.
Not this
Automation
AI replaces the investigator’s judgement.
Not this either
Augmentation
AI just adds more data for the investigator to sort through.
This
Amplified intelligence
Each partner does the part of the cognitive work it is actually best at.
Five stages, built and working today, take a case from raw evidence to an approved intelligence note.
Evidence intake
Upload CSV, Excel, Word, JSON, or PDF evidence. Every file is hashed, stored with a chain-of-custody record, and swept automatically for entities and transactions.
Ingestion summary
Files uploaded
2
Entities extracted
21
Transactions parsed
15
Chain of custody
Complete
Triage
A coverage map shows what the case has and what it is missing before any analyst time is spent, alongside a full entity inventory.
Evidence coverage
Entity resolution
Rule-based matching surfaces potential aliases and shared identifiers. Every match is a suggestion; a human confirms it, rejects it, or asks for more evidence.
Entity A
X. Xxxxx
Entity B
Xxxxxx Xxxxx
Matching: address · confidence: Medium
Typology detection
Explainable, rule-based detectors (structuring, mule networks, cross-border layering, and more) scan the evidence graph and rank leads by severity.
Structuring detected
6 transactions, $9,200 to $9,800, from one account within 72 hours.
Investigative synthesis
AI drafts an intelligence note per confirmed finding: summary, timeline, evidence gaps, jurisdictions, next steps. A human edits and approves it before it counts as anything.
Summary [AI-drafted, editable]
This is stages 1 through 5 of an 11-stage blueprint. Jurisdiction cooperation, case-decision workflows, formal report generation, supervisor review, and an agency-wide dashboard are designed but not yet built.
Entities, confirmed matches, and transaction flows accumulate into a single connected case graph as evidence comes in, not five disconnected spreadsheets. This is the real demo case's graph shape.
Every investigative decision involves three challenges, and the system allocates each differently.
Uncertainty
missing information
AI: Surfaces missing-evidence alerts and anomaly signals.
Human: Decides what the absence means.
Complexity
too many variables
AI: Detects patterns at scale across the evidence graph.
Human: Decides which lead to pursue first.
Equivocality
divergent interpretations
AI: Surfaces alternative interpretations and confidence levels.
Human: Resolves the ambiguity itself.
Every disagreement between AI and a human analyst triggers a structured divergence record, not a silent override. The human decision always stands, but the disagreement itself is never discarded. It becomes a recorded, analyzable signal, and the trigger for retuning the rule that produced it.
A disagreement rate that goes to zero and stays there is treated as a red flag, not a milestone. It usually means no one is scrutinizing the AI anymore.
These are architectural, not aspirational. They are enforced in code, not in policy documents.
No autonomous determination
AI never determines guilt or criminal classification. Every output is a lead. Every lead requires human approval.
No protected-attribute scoring
Ethnicity, religion, caste, and similar attributes are never accepted as inputs to the rule engine.
No unauthorised data access
The system only processes evidence uploaded through an authorised workflow, never external records it reaches for itself.
No black-box scoring
Every finding decomposes into the rule, the evidence, and the threshold that triggered it. No score without a reason.
No silent dissemination
Nothing leaves the system without a named human approver and a recorded decision.
No data leakage
No case data is sent to an external API or model provider, ever.
No unauditable action
Every upload, view, edit, and approval is recorded in an append-only log.
No cognitive deskilling
AI explanations are always visible. Investigators can inspect, challenge, and override every output.
TRACE Nexus is scoped to financially motivated, borderless cybercrime: crimes in which money or value moves through traceable rails, including banking, cryptocurrency, trade, and mule accounts, across jurisdictions. Any crime that leaves this kind of trail falls within scope.
Built by Himaja Prabhala, a financial crime analyst who built her early casework in the UK tracing suspicious fund flows and reviewing transaction records for fraud. All copyright in TRACE Nexus and TRACE Risk Advisory belongs to Himaja Prabhala.
The underlying approach, deterministic rule-based detection in place of opaque scoring, and citation-backed reference lookup in place of guessable retrieval, is set out in a published paper: "TRACE: An Interactive AI and Compliance-by-Design Framework for Financial Crime Detection" (Prabhala, 2026, DOI 10.5281/zenodo.20699574), documented, citable, and open to review. The first system built from that research was TRACE Risk Advisory: an SME fraud-detection engine and an AML risk-scoring engine, deterministic and rule-based, with sanctions and PEP screening against a real watchlist. Every finding decomposes into the exact rule, evidence, and threshold that triggered it, no black-box scoring.
TRACE Nexus is the bigger workspace built next: not scoring a single case in isolation, but fusing evidence across documents, resolving entities, and building a connected case graph, with a human kept in the loop at every step and explainability never traded away for convenience.
Full professional background on LinkedIn →Does this put AI and human intelligence to their best combined use to reduce crime, while preserving due process, evidence integrity, and institutional sovereignty?
If the answer is no, the feature does not exist.