ArthaShield Technology Architecture

Typed decisions.
Not prompts.

ArthaShield makes compliance decisions with a typed decision model instead of a general-purpose LLM — on a compliance-first, privacy-by-design Apache data foundation.

See the pipeline Why not an LLM

State in. Decision out.

No general-purpose LLM in the decision path.

01 · Ingest

Capture & redact

Complaints stream in; personal data is anonymized or redacted before long-term storage.

Apache Pulsar
02 · Structure

Understand the text

Explainable NLP extracts products, channels and issues into structured state.

OpenNLP · Spark NLP
03 · Decide

Typed decision

Jev answers typed questions against that state — a label plus a calibrated confidence. Never free text.

Jev · Choice + Noul
04 · Act

Gate & route

Plain code maps the decision to an action. Low confidence goes to a human.

Deterministic code
05 · Audit

Record everything

Every decision is logged before the response; access is role-based and monitored.

Audit log · Apache Ranger

Why not just ask an LLM?

Routing a complaint is a classification, not a conversation. With a general-purpose LLM you build a prompt, parse free text back into a label, hope it's a valid one — and pay full LLM cost and latency every time.

Typed, constrained output

Jev can only return one of the options it was given. No parsing step, and no chance of an invalid label.

Calibrated confidence

Every answer carries a confidence, so code decides when to act automatically and when to escalate — no costly "second opinion" call.

Cheaper and faster

System One models are built for high-volume structured classification — exactly the workload of routing every complaint.

A narrow, auditable path

State, question, typed answer, code action, log entry. Every step can be shown to a regulator — what was decided, and why.

First decision · MVP

Complaint Risk & Routing.

One call, five questions, answered in parallel: a choice across six routing decisions, and four yes/no probabilities that become the reason codes.

Jev decides	When	Code then does
LOW_RISK	Minor issue, low amount, no repeat pattern	AUTO_CLOSE_STANDARD
MEDIUM_RISK	Some risk signals, none severe	ASSIGN_TO_SUPPORT_QUEUE
HIGH_RISK	Multiple strong risk signals together	PRIORITY_QUEUE_REVIEW
FRAUD_REVIEW	Strong combined signals of unauthorized activity	SEND_TO_FRAUD_INVESTIGATION
COMPLIANCE_REVIEW	Regulatory, disclosure, mis-selling or data-privacy issue	SEND_TO_COMPLIANCE_TEAM
HUMAN_REVIEW	Ambiguous or insufficient signals	ESCALATE_TO_HUMAN_REVIEWER

Reason codes

Each is an independent probability, included at 0.5 or above.

HIGH_AMOUNT HIGH_CUSTOMER_RISK RECENT_TRANSACTION REPEATED_PATTERN

Confidence gate

If Jev's confidence is below 0.6, code overrides the decision to HUMAN_REVIEW and adds LOW_MODEL_CONFIDENCE. Ordinary application logic — not a second model call.

Built on Apache.
Private by design.

The data layer prioritizes scalability, explainability, data minimization, and regulatory auditability using proven Apache open-source technologies.

Streaming & privacy

Apache Pulsar

Real-time ingestion of complaints and case updates. Configurable retention keeps sensitive data short-lived, and Pulsar Functions anonymize or redact personal data before long-term storage.

NLP · primary engine

Apache OpenNLP

Tokenization, sentence segmentation, part-of-speech tagging, named entities (products, channels, issues) and chunking — deterministic, explainable pipelines that integrate naturally with Java BFSI stacks.

NLP · large scale

Spark NLP

Built on Apache Spark under Apache 2.0, with transformer models such as BERT and T5 — for semantic similarity, complaint deduplication, and contextual analysis across large datasets.

Search & retrieval

Apache Lucene

Text analysis, indexing and retrieval — fast search, relevance scoring, and historical case lookup across complaints and regulatory records.

Storage & minimization

Apache Parquet

Columnar storage with Snappy/Gzip compression and dictionary/RLE encoding, so only necessary fields are stored and processed — supporting FD, GDPR, and DPDP compliance.

Access & audit

Apache Ranger

Centralized access control, authorization policies, and immutable audit logs — complaint and case data access is role-based, monitored, and regulator-ready.

Pattern detection

Apache Mahout

Distributed clustering and classification of complaints and cases across millions of records — operating on distributed, aggregated datasets so patterns are found without unnecessary exposure of raw personal data.

Privacy-by-design

Personal data is minimized, anonymized early, and retained only for justified regulatory or operational purposes.

Why Apache-first

Chosen over alternatives such as NLTK or spaCy for long-term sustainability, vendor neutrality, and regulatory acceptance.

• Apache 2.0 licensing for enterprise and government adoption
• Proven scalability in large BFSI environments
• Strong governance, auditability, and ecosystem maturity
• Alignment with regulator expectations for transparency

Roadmap

One decision at a time.

Each decision is an independent module with its own evaluation set — KYC, fraud, regulatory and agent-routing decisions can be added and rolled out one by one, without touching the ones already live.

Talk to us Case Intelligence Architecture

Typed decisions. Not prompts.