Feb 2025 – Oct 2025 · 8 mos · United States

• Built scalable, distributed inference infrastructure using NVIDIA Triton with dynamic batching on AWS ECS, increasing throughput by 4x and improving GPU utilization by 30%.
• Developed an AI assistant by finetuning LLama-3.2 8B using LoRa for instruction tuning and DPO based RLHF for preference alignment, reducing support workload greatly.
• Deployed Llama-3.2 8B into production using vLLM on A10 GPUs with generation speed of 60 tokens/sec.
• Enhanced RAG pipelines with multi-attribute vector search capabilities, improving answer relevance by 25% while reducing hallucination risks.
• Developed observability and monitoring pipelines in AWS CloudWatch to track GPU/CPU utilization, reducing cloud costs by 20% and enhancing infrastructure visibility.

Sep 2024 – Sep 2025 · 1 yr · Indianapolis, Indiana, United States · Hybrid

• Conducted a VR user study with 30+ college participants using the FilmStim dataset; collected spontaneous facial expressions and eye-tracking data in VR through Meta Quest Pro’s 63 built-in Action Units aligned with FACS.
• Automated FACS data segmentation, statistical analysis, and visualisation (heatmaps, AU activation bar charts) producing robust insights into emotion intensity patterns across users.
• Built deep learning pipeline for emotion recognition through transfer learning with MobileNetV2 on unbalanced multi-class VR avatar dataset, achieving a 54.12% test accuracy.

Jul 2022 – Jul 2023 · 1 yr · Bangalore Urban, Karnataka, India · On-site

• Developed intelligent automation pipelines using Python, UiPath OCR, and Scikit-learn to extract handwritten cheque data and assess intern evaluations in a centralized, scalable platform.
• Engineered layout-aware parsing and fallback mechanisms using Tesseract OCR and OpenCV to handle varied cheque formats and complex handwriting patterns in financial datasets.
• Designed automated web-based assessment dashboards in PowerApps integrated with Power Automate and Excel VBA, streamlining intern performance tracking and feedback loops.
• Orchestrated real-time data flow using Flask, SharePoint, and SQL Server, ensuring secure capture of feedback, task completions, and evaluation metrics for HR reporting.
• Refined confidence scoring algorithms and classification models using Scikit-learn to minimize low-confidence extractions and reduce manual interventions in cheque validation workflows.
• Visualized intern progression and OCR model performance trends using Matplotlib, Seaborn, and Power BI, enabling insights into recurring issues and skill development areas.
• Facilitated continuous improvement through scheduled error analysis, model re-training, and workflow tuning using Azure Blob Storage, enhancing both HR and banking automation efficiency.

Jan 2021 – Jun 2022 · 1 yr 5 mos · India

• Constructed predictive models using Prophet, LSTM, and TensorFlow to forecast cloud resource demand on HxFlexCloud, enabling proactive auto-scaling based on usage trends and system load history.
• Orchestrated end-to-end ML pipelines in Python, automating data preprocessing, model training, and validation on time-series CPU and RAM metrics stored in InfluxDB, with performance visualizations in Grafana.
• Deployed real-time prediction services using Flask and Docker, integrated with the HxFlexCloud engine to deliver scaling recommendations and prevent over-provisioning during low-demand periods.
• Spearheaded anomaly detection mechanisms in HxSecureVault using Isolation Forest and One-Class SVM from Scikit-learn, targeting abnormal access patterns and unauthorized user behaviors.
• Streamed live access logs using Kafka and analyzed them with machine learning models to detect high-risk login attempts, leveraging ELK Stack for alert dashboards and pattern interpretation.
• Automated response workflows by connecting detection logic to AWS Lambda, ensuring timely alerts to administrators during suspicious activity events with near real-time feedback loops.
• Scheduled model retraining and deployment via Apache Airflow, enabling continuous learning on updated server logs and user behaviour profiles while maintaining accuracy and reducing false positives.