Hire Chigurupati P. - Fine-Tuning Engineer

Worked within the Apple Intelligence personalization track, supporting ML workflows that fed into features such as Siri intelligence, on-device writing assistance in Mail and Messages, and visual understanding across the iOS and macOS experiences. Engineered end-to-end Generative AI systems using Python, LLM pipelines, RAG, prompt engineering, and evaluation frameworks, improving reasoning accuracy by 12% and reducing hallucinations by 18%. Fine-tuned Mistral Transformers using PyTorch, Hugging Face, and Core ML Tools; applied distillation and pruning for 35% compression with 87%+ accuracy on-device. Developed privacy-preserving Federated Learning pipelines using TensorFlow Federated and DP-SGD Differential Privacy, enabling secure distributed training and personalized intelligence across edge devices. Built AI Agent workflows using LLM reasoning, tool calling, and workflow orchestration, integrating internal private cloud infrastructure and custom Kubernetes orchestration for highly secure, scalable AI deployment and monitoring. Automated MLOps/LLMOps workflows with Kubeflow, MLflow, Docker, and GitHub Actions, enabling CI/CD, experiment tracking, model lifecycle management, monitoring, and reducing release cycles by 30%. Collaborated with Data Science, iOS, and Platform Engineering teams to drive feature engineering, model validation, and hyperparameter tuning, increasing development efficiency by 25% and reducing preprocessing overhead by 40%.

Contributed to Stripe Radar's fraud detection stack for card-not-present payments, spanning feature engineering, model development, and real-time risk scoring used by Radar for Fraud Teams across global online checkout flows. Designed an end-to-end fraud detection ML pipeline for Stripe Radar, processing 150M+ monthly transactions and transforming payment events into scalable ML datasets using Python, Apache Spark, and SQL. Engineered 200+ behavioral, transactional, and network features including velocity signals, device fingerprinting, card history, and merchant risk indicators, improving fraud recall by 18% while controlling false positives. Developed and optimized XGBoost classification models using feature engineering, feature selection, hyperparameter tuning, cross-validation, and model evaluation, improving automated risk scoring accuracy. Automated MLOps workflows for model training, validation, retraining, and deployment, reducing iteration cycles from 2 weeks to 3 days and accelerating fraud model improvements. Deployed a real-time fraud scoring service using Scala, Apache Kafka, and distributed streaming architecture, delivering sub-100ms risk predictions at production scale. Built ML monitoring and observability dashboards tracking data drift, precision, recall, and model degradation, reducing rollbacks by 30%, executing A/B testing that lowered false declines by 12% for high-volume merchants.

Built query understanding, ranking models, and Solar-backed retrieval within Flipkart's Search and Relevance function, powering the type-ahead and search experience for the main e-commerce app and Big Billion Days traffic. Developed an NLP-based query understanding pipeline using NLTK, Word2Vec embeddings, and text preprocessing techniques for semantic matching, synonym expansion, and intent classification, reducing zero-result queries by 15%. Investigated 3M+ daily search query logs using SQL and Pandas to identify user intent gaps, query abandonment patterns, and zero-result search issues, improving the effectiveness of search relevance system for mobile users. Engineered scalable ranking features from 50M+ SKU records by transforming click-through rate (CTR), dwell time, and conversion behavior into model inputs for a learning-to-rank recommendation and information retrieval framework. Optimized an XGBoost-based ranking model with Scikit-learn through feature selection, model tuning, and offline validation, improving NDCG by 11% and precision@10 by 8% compared with the existing TF-IDF retrieval approach. Productionized the ML ranking pipeline by integrating outputs with Apache Solr search infrastructure, enabling scalable real-time search inference and reducing search response latency by 20% during Big Billion Days traffic spikes. Executed A/B experimentation across 4 city clusters using Statsmodels to validate model performance, measure ranking effectiveness, and achieve a 7% uplift in search-to-purchase conversion before full-scale production deployment.