ANUJ
DATA×AI×ENGINEERING
INITIALIZING SYSTEM
ANUJ MUNDU
RETURN TO SELECTED WORK
SOURCE CODELIVE DEMO
PROJECT 05•DATA ANALYTICS•ENGINEERING•FLAGSHIP CAPSTONE

AutoRecon Agentic™ — Financial Reconciliation & Audit

Autonomous financial reconciliation engine with agentic dispute generation, fuzzy invoice matching & conversational copilot

Python 3.11+StreamlitPandasPyPDFOpenPyXLTheFuzzRichPydantic v2FinTech Automation
Hours Saved
25+ hrs/wk
Automated AP invoice reconciliation
Fuzzy Matcher
Token-Sort
Levenshtein distance with alias mapping
Test Coverage
100%
5/5 Pytest suites passing in CI
Dispute Tones
3 Modes
Inquiry, Correction, and Escalation notices
// INTERACTIVE SYSTEM TELEMETRY & DIAGNOSTIC LAB
INTERACTIVE DECISION THRESHOLD CALIBRATOR (τ)

Slide the classification cut-off threshold to evaluate precision vs recall trade-offs and net ROI.

OPTIMAL τ:0.42 (Brier Calibrated)
DECISION BOUNDARY (THRESHOLD τ):0.42
0.15 (High Sensitivity / Catch All)0.80 (High Specificity / Conservative)
Precision
80.8%
Targeting accuracy
Recall
81.2%
Churners captured
F1 Score
81%
Harmonic mean
Net Monthly Value
+$84,723
Revenue preserved
At τ = 0.42: Caught 349 of 430 churners; 147 false alarm outreaches.
Model: XGBoost + Isotonic CalibratedCV
RUNTIME ENGINE & LATENCY BENCHMARK COMPARATOR

Empirical benchmark comparing INT8 Post-Training Quantized ONNX against vanilla TorchScript C++ tracing.

INFERENCE BATCH SIZE:
P95 Latency
24.8ms
Deterministic SLA
Throughput
40.3 FPS
Video streaming limit
RAM Footprint
14.2 MB
Model weight & graph
CPU Usage
38%
8-Core Edge node
Target: Sub-30ms budget on edge hardware✓ 3.1x Faster Than TorchScript

01 // SYSTEM OVERVIEW

AutoRecon Agentic™ is an autonomous financial reconciliation, invoice discrepancy resolution, and conversational audit copilot. Built with Python 3.11, Streamlit, and Pydantic v2, it bridges bank settlement statements with vendor accounts payable ledgers, performs deterministic and fuzzy token-sort matching, and generates ready-to-send dispute letters while offering an interactive English query copilot.

02 // THE PROBLEM & ENGINEERING SIGNIFICANCE

The Core Challenge

Accounting and finance teams spend tens of hours every week manually cross-referencing thousands of vendor invoice numbers, dates, and amounts against messy bank statement lines.

Why This Matters

Typographical discrepancies, missing reference codes, and billing overcharges slip through manual audits, leading to vendor overpayments and delayed month-end closes.

Key Constraints:
  • •Extracting structured transaction tables from varied layout PDF invoices and multi-tab Excel files.
  • •Matching line items when vendor names or invoice numbers contain typographical errors, acronyms, or alternate formatting.
  • •Drafting professional, legally defensible discrepancy dispute notices without requiring manual paralegal review.

03 // DATA PIPELINE & PREPROCESSING

Input Format: Vendor invoices (PDF, XLSX, CSV), bank statements, and ERP export ledgersSample Volume: Multi-vendor accounts payable enterprise reconciliation batches
Transformation Steps:
  • PyPDF text extraction and table boundary segmentation
  • Synonym column alias mapping (e.g. 'Inv_No', 'Bill #', 'Reference' -> standard 'invoice_id')
  • Pydantic v2 type coercion and ISO date standardization
Cleaning Strategy: Automated removal of currency symbols, whitespace trimming, and duplicate payment identification.

04 // SYSTEM ARCHITECTURE & DATA FLOW

Document Upload -> Ingestion & Synonym Normalizer -> Deterministic + Fuzzy Matcher -> Discrepancy Classifier -> Agentic Dispute Generator -> Audit Copilot UI.

STEP 01PyPDF · OpenPyXL · Pandas
Document Ingestion Layer

Parses PDF, Excel, and CSV inputs with automated schema synonym alignment.

STEP 02TheFuzz · C-Extensions
Fuzzy Matching Engine

Executes exact reference matching followed by Levenshtein token-sort ratio comparison.

STEP 03Python Logic
Discrepancy Classifier

Categorizes reconciliation states: Perfect Match, Amount Mismatch, Date Discrepancy, Orphan Invoice.

STEP 04Agentic Templates
Agentic Dispute Generator

Autonomously drafts structured dispute letters with formal legal references across 3 selectable tones.

ENGINEERING ITERATION & REFACTORING CHRONOLOGYv0 Prototype → v2 Cloud Production
v0 // PROOF OF CONCEPT

Monolithic CLI Python script with static exact string matching; failed on transposed vendor names or slight invoice number padding.

v1 // MODULARIZATION & REFACTOR

Refactored into modular Pydantic v2 data models with Levenshtein token-sort fuzzy logic and automated discrepancy categorization.

v2 // PRODUCTION CLOUD SYSTEM

Production Streamlit web platform with automated dispute letter generation across 3 tones, interactive conversational copilot, and multi-format parser pipelines.

05 // MODEL ENGINEERING & HYPERPARAMETERS

Base Architecture: Levenshtein Token-Sort Fuzzy Matcher + Deterministic Rule Engine

Benchmarked against historical accounts payable reconciliation logs with intentional typos.

Hyperparameters & Training Dynamics:
  • • Fuzzy Threshold: 85%
  • • Token Ratio Mode: Token_Sort
  • • Amount Tolerance: $0.05
  • • Date Window: +/- 5 Days
Loss Function: N/A (Fuzzy String Distance & Financial Rules)
Trade-off Rationale: Combined strict exact matching on IDs with fuzzy fallback on company names to avoid false-positive invoice collisions.

06 // FAILURE ANALYSIS & ZERO-TRUST SAFEGUARDS

OBSERVED FAILURE MODES UNDER STRESS
  • • Scanned image-only PDFs without embedded text layers.
  • • Split payments where a single invoice is paid across multiple bank statement entries.
Mitigation & Fallback: Prompts user to trigger OCR fallback on scanned documents; aggregates multi-line payments matching invoice totals.
ENGINEERING POST-MORTEM & CONSTRAINT MITIGATION
Root-Cause Analysis & Production Telemetry
Operational Constraint:

Parsing corrupted bank PDF statements without OCR latency breaking the interactive UI.

Bottleneck Encountered:

Pure OCR pipelines took 8+ seconds per page and failed on table column boundaries.

Architectural Solution:

Implemented a hybrid text-layer extraction fallback with PyPDF + bounding coordinate table heuristic prior to OCR.

Empirical Outcome:

Parsing speed increased by 5.3x with 0% data drop on digital PDFs and clean error quarantine for scanned sheets.

07 // PRODUCTION DEPLOYMENT SPECS

Serving Framework
Streamlit Cloud Interactive Web App
Containerization
Reproducible Python 3.11 virtual environment with GitHub Actions CI
P95 SLA
< 1.5s for 1,000 line item multi-file reconciliation
Throughput
Processes 600 invoice lines/sec

08 // ARCHITECTURAL DECISIONS & TRADE-OFFS

Adopted Levenshtein token-sort fuzzy matching.
Why: Token-sort handles word re-orderings (e.g., 'Acme Corp LLC' vs 'LLC Acme Corp') that cause standard edit distance algorithms to fail.
Alternative Discarded: Standard Levenshtein ratio.
Engineered 3-tone dispute notice templates (Inquiry, Correction, Escalation).
Why: Gives finance teams the flexibility to preserve gentle vendor relationships or apply immediate contractual pressure.
Alternative Discarded: Single generic email template.

09 // PLANNED IMPROVEMENTS & NEXT REVISIONS

  • →Add automated two-way ERP integration (NetSuite / QuickBooks API) for one-click invoice adjustments.
  • →Integrate vision-LLM document parsing for heavily degraded physical paper receipts.
OPEN REPRODUCIBILITY & TEST SUITE COMMAND:pytest tests/ -v --durations=5

100% pass across 5 test suites validating deterministic matcher, fuzzy tolerance, and dispute generation.

TECHNICAL INTERVIEW DISCUSSION PROMPTS
  • Q1:"Why use Levenshtein token-sort ratio instead of simple Levenshtein distance for vendor matching?"
  • Q2:"How do you handle currency conversion differences and penny rounding discrepancies across accounting ledgers?"
  • Q3:"How did you structure the 3-tone dispute engine to ensure contractual compliance without alienating vendors?"
BACK TO ALL PROJECTS
INSPECT ON GITHUB