An AI/ML Engineer working towards building Responsible and Efficient AI systems.
I believe AI can absorb a lot of mundane work and allow us to focus on complex problems solving, but that promise only holds true if we can audit what these systems produce at every step and understand the consequences of their outputs. Recently, my team and I won the Red Hat and Open Accelerator Hackathon where we built a MCP governance system that inspects incoming request. That pushed me further, and I built a prompt injection detecting system using NLP techniques.
I bring a strong technical foundation, a security-first mindset, and experience handling complex data at scale. Along with my work in healthcare CRM, I have gained exposure to research and problem-solving through a research internship at ISRO (Indian Space Research Organisation). I regularly read technical blogs and podcasts to stay current with the evolving landscape of AI and data systems. To document and reflect on my learning, I maintain a Substack where I share insights and practical takeaways .
Built and maintained ETL pipelines across 80+ healthcare data sources for pharmaceutical clients. Worked customer unification and data profiling using SQL and Salesforce Data Cloud for identity resolution across CRM lifecycle, led stakeholder walkthroughs, and owned data quality documentation. Resolved data ingestion issues and reduced storage and compute costs by 30%.
Research internship focused on CNN-based semantic segmentation for surgical instrument detection. Compared deep learning architectures across model evaluation metrics F1-score, precision and recall and visualization to identify performance on detecting instrument edges and overlapping regions.
Build a prototype to automate Variable Power Combiner tuning via MATLAB–Arduino stepper motor control. Implemented sensor feedback loops for gain stabilization in ring resonator systems. Analyzed signal variance with aim to improving precision and control while calibration in satellite communication.
Most Innovative Use of MCP at Red Hat × IBM Hackathon
A MCP Firewall which sits between external AI Agent and MCP server. Reverse proxy using LangGraph to process JSON-based requests through policy enforcement rules, validating intent and returning real-time ALLOW/BLOCK decisions.
Built an evaluation pipeline to quantify racial bias across Gemini and Ollama/Qwen models. Leveraged statistical analysis such Welch's t-test, Cohen's d; along with NLP techniques to identify bias across scores and rationale.
Built a classification system using Sentence-BERT embeddings, ChromaDB, and distance-weighted KNN. Implemented cosine similarity search over ChromaDB vector database to classify prompts as benign or malicious and t-SNE and UMAP visualizations to understand vector embeddings.
Flipkart GRID 3.0: National Finalist - Top 10 India
Developed and validated a vision-based multi-robot navigation pipeline. Used overhead cameras for perception, pose estimation, and global path planning. Implemented ArUco markers for object detection, localisation, and coordination movement.
Evaluated five deep learning architectures (ResNet-50, EfficientNet-B0, DenseNet-121, MobileNetV2, ViT-B/16) on multi-class classification of rare neurological diseases from brain MRI scans using PyTorch. Applied transfer learning with ImageNet on limited medical imaging data ( 280 samples/class) using data augmentation to reduce overfitting. Visualised CNN performance on classification metrics and Grad-Cam visualisation.
I write on Substack when an idea won't leave me alone. Some of those ideas end up getting published by Springer.
Subscribe on Substack →PRMs, ORMs, Best-of-N, MCTS, and why NVIDIA NVLink infrastructure matters for where test-time compute is going.
AI is everywhere, but understanding how the pieces fit together is hard. A visual map that turns overwhelming jargon into a coherent picture.
A Springer-published paper on using MQTT for real-time coordination between a central system and a swarm of mobile robots. Central image processing drives navigation and sends signals over a local Wi-Fi network.