AION Sentinel
Bias-Free Corporate Integrity Score. Trained on the Companies Act 2013 — not on fraud labels.
Frozen v1 234-feature manifest 6-layer anomaly engine Daily 03:00 IST pipeline
The Bias-Free Principle
Sentinel was trained on the Companies Act 2013. It does not know what fraud looks like. It knows what lawful financial statements look like — and it flags when the math does not add up.
This is a deliberate design decision. Models trained on known fraud cases learn to recognize past fraud. Sentinel learns to recognize accounting integrity — and flags anything that deviates from it, whether or not the deviation appears in any historical fraud database.
No fraud labels were used in training. No company was labelled as fraudulent or clean before the model learned what clean looks like. Sentinel derives its judgment from mathematical consistency with Companies Act obligations — not from retrospective labelling of known bad actors.
Use Cases by Stakeholder
| Stakeholder | How Sentinel Serves Them |
|---|---|
| Hedge Funds | Pre-position due diligence screening; flags accounting anomalies before corporate events |
| NBFCs | Corporate borrower integrity assessment ahead of loan disbursement |
| Mutual Funds | Portfolio monitoring for emerging filing anomalies; early-warning layer |
| Fintech Lending | Automated corporate creditworthiness signal for underwriting pipelines |
| Insurance | Commercial policyholder financial health screening |
| Regulatory | Population-level anomaly radar across listed companies |
| Retail Investors | Plain-language integrity context on companies in their watchlist |
Architecture — Six-Layer Anomaly Engine
| Layer | What It Does |
|---|---|
| Tier 1 — Data Ingestion | BSE/NSE quarterly and annual filings ingested and normalized to Companies Act structure |
| Tier 2 — Feature Engineering | 234 features derived from ratio analysis, inter-statement consistency checks, and temporal drift signals |
| Tier 3 — Frozen v1 Model | Core model layer trained on Companies Act-compliant filings; frozen to prevent drift from new inputs |
| Beneish Layer | Eight-variable Beneish M-Score replication adapted for Indian GAAP; runs in parallel with proprietary model |
| Fusion Layer | Convergent scoring: agreement between model and Beneish raises confidence; divergence is flagged separately |
| Black-Hole Memory | Persistent anomaly memory tracks whether a company has previously triggered elevated flags across filing cycles |
Classification Thresholds
| Flags Triggered | Classification | Meaning |
|---|---|---|
| 0–1 | ROUTINE | Filing consistent with Companies Act expectations |
| 2 | REVIEW | Minor anomalies; warrant a second look but are not indicative of structural issues |
| 3–4 | ELEVATED | Multiple convergent anomalies detected; higher-priority review recommended |
| 5+ | CRITICAL | Broad pattern of accounting inconsistencies; full filing audit warranted |
Sample Output — Gensol Engineering
The Gensol test case is the most publicly documented validation of Sentinel's early-warning capability.
{
"company": "Gensol Engineering Ltd",
"cin": "U40100GJ2012PLC070920",
"filing_period": "FY2024-Q3",
"risk_score": 0.1260,
"risk_percentile": 97.37,
"flags_triggered": 6,
"classification": "CRITICAL",
"anomaly_summary": "Convergent anomalies detected across receivables growth, cash conversion cycle, and inter-statement consistency. Beneish DSRI and AQI both elevated. Black-hole memory: elevated flags in prior two cycles.",
"beneish_m_score": -1.82,
"model_confidence": 0.94
}
The risk_score is a normalized output (lower = cleaner). The risk_percentile positions the company relative to the full universe of 2,435 rated companies. CRITICAL classification with flags_triggered of 5+ means the model saw broad, convergent accounting irregularities — not a single outlier ratio.
Public Risk Radar
Sentinel publishes a daily snapshot of companies currently in ELEVATED or CRITICAL classification. Available as a structured JSON feed at /sentinel/risk-radar.json. No account required for the radar feed.
Infrastructure
In development: infrastructure to support financial-regulatory integrity.
Limitations
- Sentinel scores financial statement consistency, not management intent. A CRITICAL score means the filings are mathematically irregular — it does not mean fraud has occurred.
- The model is frozen at v1. New accounting irregularity patterns that emerge post-training will not retroactively improve the model's recall on novel schemes.
- Coverage is limited to listed Indian companies with structured BSE/NSE filings. Unlisted entities are not in scope.
- Small-cap companies with limited filing history have lower confidence scores — the Black-Hole Memory layer cannot operate without multi-cycle data.
- IFRS-converted filings require additional normalization; some precision is lost in the mapping to Indian GAAP assumptions.
- This is not SEBI-registered advisory output. Sentinel scores are analytical context, not regulatory determinations.