AI / DATA / RESEARCH

Forensys

Anomaly detection and decision support for tax compliance and revenue assurance.

Research and academic work, evolving into a product surface. Not a deployed commercial tax product.

Year: 2025–2026
Status: Research
Role: Designed the signal framework, risk-scoring models, evaluation methodology and analysis interface.
Categories: ResearchAnomaly DetectionFinancial SystemsMachine LearningDecision Support

Forensys company analysis interface showing NGX-listed companies with risk scores and forensic signal tabs

The problem

Tax compliance and revenue assurance depend on spotting patterns that do not add up: revenue that does not reconcile with cash flow, receivables that age oddly, inventory that never turns, profit and tax that disconnect.

Analysts cannot read every filing at depth. The problem is not lack of data — listed companies publish plenty — but lack of a systematic first pass that ranks where human attention should go, with enough explanation to trust the ranking.

Labels are scarce. Confirmed fraud cases are rare, delayed and biased, so conventional supervised classification alone is a weak fit for the domain.

Context

Research and academic work exploring anomaly detection on financial disclosures from Nigerian Exchange (NGX)-listed companies, with an interface prototype for analyst workflows.

Presented as research. It is not a commercial tax product, not certified for regulatory use, and not deployed with production clients.

My role

Designed the forensic signal framework: named, interpretable checks (revenue/cash-flow gap, receivables ageing, inventory quality, profit–tax disconnect) each with a threshold and plain-language interpretation.
Built risk-scoring models and evaluated them honestly — including distributional overlap analysis and per-class precision/recall/F1 — rather than reporting a single headline number.
Implemented the analyst interface: company lookup, signal panels, tax compliance view, network graph and ML performance tabs.
Defined evaluation methodology suited to rare-event detection: score distributions, separation distance, and class-wise metrics against a stated threshold.

The solution

A Forensys workspace where an analyst selects a listed company and reviews a risk score with drill-downs: overview, forensic signals, tax compliance, network graph and ML performance.

Interpretable signals — not black-box scores alone. Each card shows the metric, the threshold, a plain-language note (e.g. DSO in normal range) and recommended investigative steps.

Model evaluation made visible inside the product: score distributions for clean vs. fraud cohorts and per-class metrics, so users understand discrimination limits instead of trusting a single number.

Exportable reports so findings can move from the tool into human review workflows.

Artifacts

Forensys forensic signals panel showing REVENUE_CF_GAP, RECEIVABLES_AGING, INVENTORY_QUALITY and PROFIT_TAX_DISCONNECT scores against thresholds — Forensic signal cards with thresholds and plain-language notes.

Risk score distribution chart comparing clean and fraud samples with means of 22.3 and 26.9 and a separation of 4.6 points — Score distribution: clean vs. fraud — separation analysis.

Per-class classification performance bar chart for clean vs fraud showing precision, recall and F1 scores — Per-class precision, recall and F1.

Forensys dashboard interface with navigation for document analysis, company lookup, tax rules and reports — Analyst workspace.

Architecture

Forensys analysis pipeline

System flow

01Source dataPublic filings · NGX companies
02Feature pipelineRatios · gaps · ageing · tax
03Signal engineNamed checks + thresholds
04Risk modelScoring · calibration
05Analyst UILookup · panels · reports

Implementation

Feature engineering focuses on interpretable financial ratios and relationships rather than opaque embeddings — every signal must be explainable to an auditor.

A rules-plus-model hybrid: deterministic threshold checks produce explainable flags; a learned score ranks severity and surfaces non-obvious combinations.

Evaluation embraces the hard truth of rare-event detection. On the current eval set (clean n=300, fraud n=120), score distributions overlap substantially — mean separation of 4.6 points — which is reported openly rather than hidden behind accuracy claims.

Per-class metrics (e.g. fraud-class recall 0.62 and F1 0.75 at a 0.80 threshold on the current split) drive iteration priorities: recall on the minority class matters more than overall accuracy here.

The interface is an analyst tool, not a dashboard demo: company context, signal detail, investigative steps and export are the core loop.

08 / Hard parts

The hard engineering.

Almost no trusted labels

Confirmed fraud is rare, delayed and selected. Building evaluation you can defend with limited positive labels — and being honest about what the metrics mean — was harder than training the model.

Interpretability is a requirement, not a nice-to-have

A risk score nobody can explain is unusable in compliance work. Every signal was designed to carry its own evidence: metric, threshold, plain-language note and suggested checks.

Distribution overlap is real

Clean and fraud score distributions overlap heavily on current data. Rather than overclaim, the system surfaces separation analysis so users calibrate trust — a 4.6-point mean separation informs how the score should be used as a triage tool, not a verdict.

Research versus product

The work needed to stay honest as research while still producing a usable interface. That meant separating evaluation views from decision views and labelling the system's limits inside the product itself.

Results

A working research prototype: interpretable forensic signals over public financial data, risk scoring with visible evaluation, and an analyst interface covering company lookup, signal drill-down and reporting. Evaluation on the current holdout (clean n=300 / fraud n=120) shows meaningful but limited discrimination — mean score separation 4.6 points, fraud-class F1 0.75 at threshold 0.80 — reported as a triage aid under active research, not a production compliance guarantee.

4.6 pts

Mean score separation

Clean (n=300) vs fraud (n=120)

0.75

Fraud-class F1

At 0.80 threshold, current split

0.62

Fraud-class recall

Priority metric for rare events

Named forensic signals

Each with threshold + interpretation