Prefiler — Product Summary
- Category: AI accounting intelligence platform
- Primary users: Accountants, CPAs, chartered accountants, tax professionals
- Core function: Analyse financial data before tax filing
- Key capabilities: Transaction classification, risk detection, clarification questions, financial summaries
- Risk scoring: Deterministic, rule-based, materiality-weighted, version-controlled
- AI model: Used for narrative generation only — architecturally separated from scoring
- Data handling: Ephemeral processing, no permanent raw document storage
- Bank support: Profile-driven parsing for HDFC, SBI, ICICI, Axis, Kotak, PNB, Yes Bank, and more
System Architecture
- Hybrid intelligence model: deterministic rule engine + heuristic pattern detection + LLM reasoning layer
- Strict output enforcement: all AI responses conform to JSON schema and pass validation before rendering
- Deterministic risk scoring: materiality-weighted, version-controlled framework producing reproducible results
- Architectural separation: AI-generated narratives are structurally separated from deterministic scoring outputs
- Output validation pipeline: schema validation → confidence scoring → risk factor extraction → rendering
Document Extraction Pipeline
- Profile-driven deterministic parsing for digitally generated PDFs across major Indian banks
- Structured bank layout profiles: column patterns, date formats, polarity conventions, footer filtering
- Universal header matching: three-tier strategy (exact → starts-with → keyword) for column detection
- Vision AI fallback for scanned documents lacking text-layer content
- Integer cents arithmetic for reconciliation: eliminates floating-point precision errors
- Reconciliation verification: computed closing balance checked against statement closing balance
Transaction Classification
- Heuristic pattern detection: narration text analysis for transaction type identification (UPI, NEFT, IMPS, cheque)
- Category mapping: income, expenses, transfers, loan transactions, investment activity, regulatory payments
- Confidence scoring: each classification includes a confidence percentage
- Ambiguity handling: structured clarification questions surfaced when classification confidence is low
Risk Scoring Engine
- Deterministic, rule-based framework with materiality weighting
- Scoring dimensions: transaction concentration, regulatory thresholds, cross-category patterns, data consistency
- Non-linear materiality scaling with effective turnover
- Multi-signal escalation logic for compound risk indicators
- Version-controlled methodology: every analysis records the framework version used
- Explainability: every risk score is traceable to specific triggered rules and thresholds
Data Governance
- Ephemeral document processing: raw bank statements are not permanently stored
- Derived metadata only: transaction classifications, risk scores, and summaries are retained
- Sensitive field masking in logs
- Row-level security (RLS) enforcement at the database level
- Firm-scoped data isolation for multi-tenant deployments
- Role-based access control: Partner, Senior CA, Junior with corresponding permissions
Audit & Quality Infrastructure
- Audit trail: operation logging with timestamps and user attribution
- Reasoning trace logging for explainability review
- Benchmark testing: classification accuracy measurement against human-reviewed samples
- Extraction telemetry: parse confidence, reconciliation status, processing timings
- LLM model and prompt version tracking per analysis
Related Resources
- Risk Scoring Methodology — Detailed framework documentation
- Technical Trust — Architecture trust points
- Bank Statement Analysis — Extraction pipeline details
- Data & AI Policy — Governance and AI usage
- Security Overview — Infrastructure and access control