I build production ML systems that serve 10K+ RPS at <25ms, fine-tune LLMs that unlock INR 15Cr monthly, and optimize infrastructure saving $18K/month.
Machine Learning Engineer with 3+ years of experience building production ML systems at Navi Technologies. I work across the full ML lifecycle — from fine-tuning LLMs for SMS entity recognition to developing credit risk models that improved approval rates by 150bps.
I'm driven by optimizing systems and processes: reducing model training costs by 31%, saving up to $18,000 monthly through AWS resource optimization, and architecting real-time feature serving at sub-25ms p99 latency. Currently focused on MLOps infrastructure and GenAI applications in financial services.
Proposed and implemented parameter-efficient fine-tuning of Mistral 7B and LLaMA 3 8B using QLoRA for SMS entity recognition. Distilled into BiLSTM-CRF with BIO tagging to solve production latency constraints. Eliminated manual tagging dependency and reduced turnaround by 24 hours.
Built an end-to-end guided re-verification platform for rejected loan applicants. Integrated in-house ITR & GST parser with a multi-agent LLM engine and multi-lingual chatbot. Orchestrated a dynamic 3-5 model ensemble generating 150+ features.
Replaced XGBoost with CatBoost for PrismV6 risk model, engineering new features for a direct 150bps approval rate increase. Built income models using XGBoost + LightGBM improving accuracy within 10% by 6 points.
Architected a family of 4 services serving SMS, app, device and location features in real-time. Leveraged Reactive Kotlin with Postgres, Scylla, EFS, S3, and Kafka. Implemented Trie-based template searching to cut runtimes by 70%.
Designed a comprehensive MLOps platform with FastAPI backend, PostgreSQL, MLflow integration for model versioning. Automated pipeline switching between prod environments with JIRA & Slack notifications.
Built Databricks access control framework and AWS resource usage regulator. Analyzed compute patterns and built a model to recommend resource reductions. Optimized feature-selection jobs (RFE/IV/Boruta) using Ray and tensors.
Attention model generating production-ready code from natural language and wireframe diagrams. Custom token vectors support multi-framework translation (React, Angular).
IoT device using ESP32 + Arduino Nano capturing images over WiFi with real-time scene descriptions via YOLOv3 and Transformer model, trained on MS COCO.
Image captioning model with 1.5M+ parameters combining LSTMs and CNNs, trained on Flickr30k for automatic description generation.
Utilized ResNet50 to extract features and distinguish COVID-19 from normal lung X-rays and pneumonia, achieving 99% accuracy on diagnostic imaging.
CNN-based model to classify skin lesions into Benign and Malignant with 86%+ accuracy using XceptionNet for segmentation and custom Dense Network.
I'm always open to discussing ML engineering roles, interesting projects, or opportunities to collaborate. Let's build something remarkable.
kartikeya72001@gmail.com