
Associate Machine Learning Engineer
Jocata Financial Advisory and Technology · Jun 2024 – Present
- Developed end-to-end credit risk scoring models using bank statement, credit bureau, and Days Past Due (DPD) data to predict customer creditworthiness.
- Built and optimized a scalable scoring pipeline processing 300-400 bank statements per minute, reducing end-to-end runtime by 10% and accelerating credit risk assessment.
- Developed Savings Account and Current Account scoring models, improving ROC-AUC by 14% and 13%, respectively, while increasing the Gini coefficient by 10% for the Savings Account model and 17% for the Current Account model.
- Performed model validation, testing, debugging, and production integration while optimizing data processing and scoring pipelines to ensure accurate, scalable, and reliable credit risk assessment.
- Developed and deployed a New-to-Credit (NTC) machine learning scoring model for first-time borrowers using alternative financial and transaction-based features, achieving a ROC-AUC of 0.82 and Gini coefficient of 0.64.
- Developed an end-to-end machine learning fraud-detection pipeline capable of processing approximately 200 GSTINs per minute, enabling large-scale screening of GST filings and improving audit intelligence workflows.
- Built supervised learning models to detect fraudulent filings, excessive or ineligible Input Tax Credit claims, tax misclassification, underreporting, and tax-evasion patterns, achieving a 95% fraud-detection rate while reducing false positives by 84%.
- Engineered domain-specific features from registration, return-filing, transaction, tax-payment, and Input Tax Credit data, and applied feature selection, hyperparameter tuning, and risk-scoring techniques to prioritize high-value fraud cases with greater potential revenue impact.
- Developed anomaly-detection and network-analysis frameworks that uncovered an additional 4% of fraudulent cases missed by supervised models and identified circular trading, coordinated GSTIN fraud, group-level tax evasion, and fraudulent Input Tax Credit networks.
- Developed an end-to-end LLM-powered self-service analytics platform that enabled business users to query enterprise databases conversationally without writing SQL; built a RAG-based, agentic NL-to-SQL pipeline using schema metadata, column descriptions, data types, semantic search, and primary/foreign-key relationships to select relevant tables, generate multi-table JOIN queries, execute SQL, and produce DataFrames and visualizations across tables with 80+ columns.
- Evaluated and validated the complete platform across four LLMs-LLaMA 3.1 8B, LLaMA 3.2 3B, GPT-3.5, and GPT-OSS 20B-tracking SQL accuracy, prompt iterations, model regressions, schema ambiguity, SQL operations, and edge cases to strengthen production reliability and deployment readiness.




