Shivam

Built a banking document-automation framework: a Siamese document classifier (HPO with Ray Tune, TensorRT-optimized for low-latency routing) sends docs to category-specific extraction pipelines; outputs feed an LLM layer served via vLLM, improving comprehension and response quality. Re-architected ORMB paymentinvoice reconciliation with candidate retrieval plus exact matching and vectorized transactions, cutting runtime 512s to 1.6s on 1.4M transactions and enabling horizontal scale. Developed an anomaly-detection ensemble on corporate lending data to surface risk signals across multiple models. Implemented an in-memory multivariate anomaly predictor for corporate billing to address monthly billing anomalies. Built a Slack RAG Q&A bot that answers from channel context and returns cited message references. Solved loan securitization and optimized investment allocation using linear constraint optimization.

Built and shipped a flight-ticket extraction service (regex, NER, rules) with human-in-the-loop; deployed on GCP Cloud Run as a reliable API for a travel-assistant recommender, achieving 0.92 recall in production.

Built an auto-tag metadata extraction app for document search using PositionRank utilizing font features; containerized on IBM OpenShift and integrated the API with Power Automate, delivering mAP 0.73 and faster lookup for sales. Delivered a healthcare PII detection solution with BERT fine-tuning, reaching 0.89 recall and supporting GDPR/CCPA compliance. Shipped a predictive defect analytics tool with semantic search over JIRA/ALM histories, reducing repetitive query tickets by 70%.