Project Anant

⚡

2M Transactions in ~3.8s

Complete SIMD-accelerated DuckDB OLAP pipeline, executing parallel CSV ingestion, feature extraction, and multi-signal scoring an order of magnitude faster than the 60-second baseline mandate.

🛡️

7-Signal Dual-Stage AML Model

Calibrated Bayesian Noisy-OR inference fusing turnover conservation, temporal velocity, crypto terminal exits, and foreign IP/bot fingerprints. 100% recall on 1,073 mules; 0% false positives on 23,500 citizens.

🕸️

122 Syndicates Clustered

Weighted Label Propagation and community modularity clustering 1,073 mules into 122 organized crime syndicates with automated Layer 0–3 hierarchy and 4-hop money flow tracing.

⚖️

Court-Ready Legal AI

Section 91 Cr.P.C. / Section 94 BNSS bank freeze notices and Section 172 Cr.P.C. Police Case Diaries in bilingual LaTeX with official law enforcement watermark (28% opacity).

🚀

Powered by Aegon C++26

Linux-native io_uring asynchronous engine, stepped SIMD vectorization (AVX-512/AVX2), HTTP/2 multiplexing, and nanosecond static asset cache.

🔒

100% Offline & Private

Zero external cloud reliance. Local embedded DuckDB, SIMD memory graph, and local LLM (Ollama Gemma) ensuring sensitive banking records never leave on-premise infrastructure.

Key Performance Benchmarks

Metric	Standard Baseline Mandate	Project Anant Achieved	Advantage
Ingestion & Scoring Speed	$< 60$ seconds	~3.84 seconds	15.6× Faster
Dataset Scale	2,000,000 transactions	2,000,000 transactions	100% Full Dataset
Unique Accounts Processed	24,873 accounts	24,873 accounts	Complete Graph Coverage
Mule Recall Rate	$> 90 %$ expected	100.0% (1,073 / 1,073)	Perfect Fraud Recall
Clean Citizen False Positives	Low	0.0% (0 / 23,500)	Zero Innocent Citizens Frozen
Victim Account Protection	Standard Models: None	100% (300 / 300 shielded at L0)	Automated Shielding ( $\leq 8.5$ )
Syndicate Identification	Unsupervised	122 distinct rings clustered	Label Propagation Graph Mining
RAM Utilization	Standard hardware ( $< 16$ GB)	$< 1.5$ GB peak memory	C++ Zero-Copy SIMD Efficiency

Metric

Standard Baseline Mandate

Project Anant Achieved

Advantage

Ingestion & Scoring Speed

< 60

seconds

~3.84 seconds

15.6× Faster

Dataset Scale

2,000,000 transactions

2,000,000 transactions

100% Full Dataset

Unique Accounts Processed

24,873 accounts

24,873 accounts

Complete Graph Coverage

Mule Recall Rate

> 90 %

expected

100.0% (1,073 / 1,073)

Perfect Fraud Recall

Clean Citizen False Positives

Low

0.0% (0 / 23,500)

Zero Innocent Citizens Frozen

Victim Account Protection

Standard Models: None

100% (300 / 300 shielded at L0)

Automated Shielding (

\leq 8.5

)

Syndicate Identification

Unsupervised

122 distinct rings clustered

Label Propagation Graph Mining

RAM Utilization

Standard hardware (

< 16

GB)

$< 1.5$ GB peak memory

C++ Zero-Copy SIMD Efficiency

Architectural Workflow

[Raw Banking CSV (2M Rows)]
          │
          ▼
[DuckDB SIMD Parallel Ingestion] ──── (2.7s)
          │
          ▼
[MuleScorer: 7-Signal Noisy-OR] ───── (1.0s) ───► [Dual-Stage Gate: Fraud vs Clean vs Victim]
          │
          ▼
[Syndicate Graph Clustering (LPA)] ── (70ms) ───► [122 Crime Rings & L0-L3 Layers]
          │
          ▼
[Aegon C++26 HTTP/2 Server Core] ─── (Sub-ms) ──► [REST & SSE Telemetry APIs]
          │
          ▼
[React 18 WebGL Sigma.js Studio] ───────────────► [Interactive Forensics & Case Diary PDF]

Project AnantAML Forensic Intelligence & Graph Engine