Deep Learning & Computer Vision

• CNN + LSTM real-time anomaly detection on AV telemetry (IMU, CAN, GPS)
• Transfer learning pipelines — ResNet50, VGG16, DenseNet121 — for species and flare classification
• Spatio-temporal models with GRU + Attention for satellite image sequences
• Generative model benchmarking: Conditional DCGAN vs VAE vs Diffusion Models on imbalanced astronomical data

LLMs · GenAI · LLMOps

• Enterprise LLMOps platform with automated regression detection and CI/CD quality gates
• RAG pipelines with LangChain for low-latency intelligent Q&A (62% latency reduction)
• Multi-Agent RL (DQN-based MARL) for autonomous vehicle cybersecurity
• NLP sentiment analysis, chatbot development, and prompt engineering workflows

Data Science & Machine Learning

• End-to-end ML pipelines: EDA → feature engineering → model training → deployment
• Market intelligence platforms: K-Means clustering, Random Forest, predictive analytics
• KPI reporting, statistical modelling, and business insight extraction at scale

Deployment & BI

• Production REST APIs (Flask / FastAPI) at 100+ req/sec, sub-200ms latency
• Interactive Streamlit dashboards with recommendation systems and business advisors
• Power BI and Tableau dashboards: KPI cards, regional heatmaps, dynamic slicers
• Client-side ML deployment via TensorFlow.js — zero backend infrastructure

AI / ML Engineer Intern — TechZia   Jan 2026 – Jun 2026

Autonomous Vehicle Cybersecurity · Bengaluru

• Architected a CNN + LSTM real-time anomaly detection pipeline on AV telemetry streams (IMU, CAN bus, GPS) within ROS 2 Humble, running continuous inference across 3 concurrent sensor channels
• Engineered a DQN-based Multi-Agent RL (MARL) framework with 3 specialised agents — sensor security, vehicle behaviour, network cybersecurity — orchestrated by a hierarchical global decision agent
• Preprocessed the comma.ai AV dataset via vectorised NumPy pipelines; researched CNN-Transformer hybrid and GNN-based multi-sensor fusion architectures for vehicle network intrusion detection

Data Science Intern — Seeroo IT Solutions   Apr 2025 – May 2025

NLP Pipelines · Conversational AI · Bengaluru

• Optimised NLP inference pipelines achieving 62% latency reduction while sustaining 500+ queries/day throughput under production load
• Built and deployed an end-to-end AI chatbot using LangChain and RAG architecture for domain-specific intelligent Q&A

Model Regression Detection System

Python Streamlit Groq Llama 3.1 GitHub Actions

Enterprise-grade LLMOps platform for evaluating LLM prompt versions, automatically detecting performance regressions, and generating deployment-ready evaluation reports — integrated into CI/CD via GitHub Actions.

• Benchmarks prompts against golden datasets; tracks classification accuracy across all historical versions
• Automated quality gate: blocks deployment on detected regression
• Interactive dashboard — leaderboards, regression history, performance trends
• Downloadable reports in JSON, CSV, and HTML; modular multi-provider architecture

Zomato Market Intelligence Platform

Python Scikit-learn Plotly Streamlit TextBlob

End-to-end data science + BI platform analysing Bengaluru's restaurant ecosystem across 12,000+ listings — from raw data to a live business advisory tool.

• Full pipeline: data cleaning → EDA → feature engineering → sentiment analysis → K-Means segmentation → Random Forest rating prediction
• Market Expansion Advisor — scores locations by cuisine demand, competition density, budget, and a custom success metric
• Streamlit dashboard: market overview, location intelligence, cuisine demand maps

Automated Species Identification

Python ResNet50 VGG16 TensorFlow Flask

CNN transfer learning on 10,000+ images across 90 bird species. Accuracy lifted 35% → 91.76% through systematic hyperparameter tuning (learning rate, batch size, dropout).

• Production REST API (Flask) — 100+ req/sec, <200ms inference latency

Solar Flare Classification

Python DenseNet121 GRU Attention TensorFlow.js React

Spatio-temporal model on 8,000+ satellite image sequences — 75% overall accuracy, 58% M-class recall, 36% X-class rare event recall.

• Deployed client-side via TensorFlow.js in React — zero backend infrastructure required
• Research published at IEEE ICVADV-2026

Retail Profitability Intelligence Dashboard

Python SQL Power BI Pandas NumPy

Engineered a custom Profit Margin column and Risk Score metric to flag loss-making transactions from a retail dataset. Built an interactive Power BI dashboard — KPI cards (Total Sales, Profit, Orders, Avg Margin), category-wise sales, regional profit heatmaps, and dynamic slicers — to surface actionable pricing and inventory insights via SQL aggregations.

IEEE-Indexed Research Paper   Feb 2026

Synthetic Data Augmentation for Robust Solar Flare Classification: Comparative Analysis of Conditional DCGAN, VAE and Diffusion Models

IEEE ICVADV-2026 · IEEE Computational Intelligence Society

Benchmarked three generative architectures for synthetic astronomical image generation on severely class-imbalanced datasets. Demonstrated measurable improvements in downstream classifier robustness across M-class and X-class solar flare events.

Government of India Design Patent   Jan 2025

Mobile Cooler Class 23-04 · Patent No. 443142-001

Original industrial design co-developed and officially registered with the Government of India's Office of the Controller General of Patents, Designs & Trade Marks.