Our Solution

JupiterBrains delivered a domain-specific Kannada-English name matching engine powered by small language models fine-tuned on Indian name datasets, phonetic encoding algorithms designed for Dravidian languages, and hybrid rule-based plus machine learning scoring. The solution operated entirely on-premise within BBMP's infrastructure, processing millions of records with 98.4% accuracy while maintaining sub-second API response times for real-time citizen queries. 

Custom Transliteration and Phonetic Encoding 

The foundation was a sophisticated transliteration engine trained on the Antarakshar and Apsara-Antara datasets—comprehensive collections of Kannada-English name pairs covering regional variations, phonetic patterns, and transliteration conventions specific to Karnataka. The system normalized Kannada names into standardized phonetic representations, applying letter-stability rules that weighted consonants and initial characters more heavily than vowels and trailing characters, since these remain more consistent across spelling variations. 

The phonetic encoder generated stable signatures for names that captured sound patterns rather than exact spelling. For example, "Shrinivasa," "Srinivas," and "Sreenivas" all mapped to similar phonetic encodings that scored as high-confidence matches despite surface-level string differences. The encoding handled Kannada-specific phonetic features including aspirated consonants, retroflex sounds, vowel nasalization, and consonant clusters common in Sanskrit-derived names. 

For ambiguous cases not present in training data, the system employed a controlled fallback to Google Translate API, using external translation only when internal datasets lacked coverage. This hybrid approach maximized accuracy on common Karnataka names while maintaining reasonable performance on rare or newly encountered name variants.

 Hybrid Rule-Based and Machine Learning Scoring 

The matching engine combined deterministic linguistic rules with learned similarity patterns. Rule-based components encoded domain knowledge about Karnataka naming conventions: recognizing that "S. Ramesh" and "Ramesh S" likely refer to the same person, handling patronymics and village names that appear in varying positions, and identifying common abbreviation patterns (K. N. for "Karnataka Nagara," D/o for "daughter of"). 

Machine learning models trained on historical BBMP, BESCOM, and Kaveri match pairs learned subtle similarity patterns that rules couldn't capture. The ML component scored phonetic similarity, character edit distance with position-weighted penalties, n-gram overlap, and contextual signals from surrounding address or property data when available. The hybrid scoring system merged rule-based and ML confidence scores, producing a final match confidence percentage for each candidate pair. 

High-confidence matches (above 95% threshold) automatically proceeded to record linkage. Medium-confidence matches (80-95%) were flagged for human review with specific discrepancy highlights. Low-confidence matches (below 80%) were excluded from automatic processing, preventing false positives that could incorrectly merge unrelated individuals' property records. 

GAN-Driven Accuracy Refinement for Edge Cases 

Generative Adversarial Network (GAN) techniques refined the matching engine's performance on difficult cases including nicknames and shortened forms, initials versus full names, missing or transposed middle names, and names with extensive vowel variations. The GAN framework generated synthetic name variation pairs, training the discriminator to distinguish genuine matches from false positives with increasing precision. This adversarial training improved the model's ability to handle edge cases that appeared infrequently in training data but occurred regularly in production. 

Batch Processing and Real-Time API Integration 

The system supported two operational modes addressing different use cases. Overnight batch processing reconciled the complete BBMP, BESCOM, and Kaveri databases, identifying 300,000 previously unmatched property records where ownership could now be definitively established. These batch jobs processed up to 300,000 records daily, running on BBMP's CPU cluster without GPU infrastructure requirements. 

Simultaneously, a real-time API served citizen-facing e-Khata applications and ARO office workflows. When citizens submitted khata requests or officers queried property ownership, the API returned match results with confidence scores in milliseconds. The API integration enabled Assistant Revenue Officers to focus on exception handling rather than routine verification, dramatically improving throughput. 

On-Premise Deployment with Data Governance 

Recognizing the absolute requirement for data sovereignty, the entire system operated within BBMP's secure VPC (Virtual Private Cloud) with network isolation. All processing occurred on CPU-optimized servers, eliminating expensive GPU dependencies and reducing operational costs by 90% compared to GPU-based alternatives. Zero external data transmission occurred during production operations (except controlled Google Translate API calls for rare edge cases, which transmitted only name strings without associated PII context). 

Role-Based Access Control (RBAC) ensured that only authorized revenue department personnel accessed citizen name data, with all queries logged for audit purposes. Comprehensive audit trails recorded every match decision, confidence score, and human review action, providing complete traceability for regulatory compliance and dispute resolution. The governance framework aligned with India's emerging data protection regulations and Karnataka state government data handling policies. 

Deployment: Fully on-premise within BBMP data center, network-isolated VPC 
Agents Used: Himalia (Document & Data Intelligence for name parsing and matching), Sinope (Compliance & Audit Logging) 
Technologies: Antarakshar/Apsara-Antara datasets, phonetic encoding algorithms, hybrid rule-based + ML scoring, GAN-driven refinement, CPU-optimized inference 
Timeline: 6 weeks from project initiation to production deployment with e-Khata integration