VRUSHALPATIL

PTC

• ▹Architected an autonomous multi-agent orchestration layer using LangChain, enabling cross-database reasoning across Snowflake and PostgreSQL with dynamic task routing.
• ▹Engineered Redis-backed semantic caching delivering an 801× query speedup (31.9s → 0.03s), slashing redundant LLM API calls by 60–90% and serving cache hits at 10–50ms latency.
• ▹Built a hybrid RAG pipeline (semantic + keyword retrieval) and modular Model Context Protocol (MCP) servers, optimizing pipelines over 1.15M+ tokens with 4.8× faster responses.

PTC

• ▹Built a RAG-based AI agent capable of translating natural language requirements into executable Selenium test scripts, reducing manual QA engineering effort by 80%.
• ▹Optimized retrieval performance and context management using semantic caching, cutting generation latency by 3x.
• ▹Delivered an autonomous software validation pipeline where the agent iteratively generates, executes, and self-corrects testing scripts based on environment feedback.

Pune MahaMetro

• ▹Developed and trained an advanced computer vision model using YOLOv8 for real-time platform safety monitoring and automated anomaly and accident detection.
• ▹Architected an intelligent edge-inference agent to process live CCTV feeds, achieving 95% detection accuracy while dynamically triggering automated safety alerts for immediate incident response.
• ▹Containerized the complete ML pipeline via Docker for scalable edge deployment, optimizing inference speed and hardware utilization for resource-constrained environments.

Smart Infrastructure & Edge Vision

• ▹Trained an Ultralytics YOLOv8m object detector for 70 epochs on road surface imagery to identify and localize potholes and structural asphalt hazards.
• ▹Achieved real-time GPU inference speeds of 11–14 ms per frame (~75–90 FPS), enabling high-speed onboard vehicle deployment for automated road condition audits.
• ▹Validated precision and recall curves via confusion matrices and deployed end-to-end video inference across continuous road footage.

Video Analytics & Public Safety

• ▹Built an autonomous video analytics pipeline powered by a custom-trained YOLOv5 detector to classify surveillance feeds into harassment vs. safe public interactions.
• ▹Implemented frame-by-frame inference with dynamic bounding-box visual overlays, confidence thresholding, and real-time incident categorization.
• ▹Automated video compilation with OpenCV VideoWriter, exporting fully annotated incident monitoring streams for rapid security intervention.

Computer Vision & Surveillance AI

• ▹Architected a deep convolutional bounding-box regression model utilizing a DenseNet121 backbone to localize physical harassment incidents in surveillance frames.
• ▹Developed a streaming batch data generator parsing Pascal VOC XML annotations with real-time coordinate normalization and image preprocessing.
• ▹Trained utilizing Smooth L1 loss and mean Intersection over Union (IoU) evaluation, featuring a custom callback to automatically checkpoint peak validation IoU weights.

Industrial AI & Reliability Engineering

• ▹Architected an industrial predictive maintenance deep network utilizing stacked LSTM layers with dropout regularization to predict Remaining Useful Life (RUL) of aircraft jet engines.
• ▹Engineered multi-sensor time-series pipelines on NASA C-MAPSS run-to-failure telemetry using MinMax feature scaling and 60-cycle sliding-window sequence generators.
• ▹Achieved a test Mean Absolute Error (MAE) of 12.40 operational cycles and an R² score of 0.756 (75.64%) on unseen test jet engine degradation profiles.

Medical AI & Healthcare Diagnostics

• ▹Developed a non-invasive early diagnostic screening pipeline using Convolutional Neural Networks (CNN) to detect Parkinson's disease from motor tremor patterns in hand-drawn spirals and waves.
• ▹Applied real-time data augmentation (random shear, zoom, horizontal flips) via Keras ImageDataGenerator to prevent overfitting on clinical drawing samples.
• ▹Achieved 93% test classification accuracy on unseen test spiral drawings utilizing early stopping and adaptive learning rate plateau scheduling.

Bioinformatics & Computational Biology

• ▹Conducted large-scale benchmarking of 9 machine learning classification algorithms across 16,000+ protein structures and crystallographic attributes from the Protein Data Bank (PDB).
• ▹Engineered robust feature preprocessing pipelines addressing extreme class imbalance, pH outlier filtering, Matthews correlation density, and molecular weight normalization.
• ▹Achieved 87.87% accuracy with Decision Trees and 86.47% test accuracy with 10-fold cross-validated Random Forest ensembles across complex bio-molecular classes.

Operations Research & Geospatial AI

• ▹Formulated a Teaching-Learning-Based Optimization (TLBO) metaheuristic to solve the Capacitated Vehicle Routing Problem (CVRP) across distributed depots, warehouses, and delivery nodes.
• ▹Integrated Bing Maps REST API for real-time driving route matrix calculation, duration estimation, and polyline coordinate fetching under realistic road network conditions.
• ▹Enforced hard truck payload limits (4,000 kg capacity) and driver operational thresholds (9-hour maximum shift limit with 30-min per-stop delivery latency penalty).
• ▹Rendered dynamic, interactive multi-fleet GPS routes and depot markers using Folium GIS mapping for logistics visualization.

Springer LNNS / Vishwakarma Institute of Technology

• ▹Engineered an end-to-end NLP classification pipeline benchmarking 9 machine learning algorithms across 47,000+ multi-class social media and Twitter/X records.
• ▹Formulated advanced text preprocessing using tokenization, lemmatization, stop-word filtering, and TF-IDF feature extraction to capture nuanced linguistic patterns.
• ▹Achieved 92.92% accuracy and 94.5% F1-score with an optimized Ensemble Bagging Classifier, outperforming standard baseline classification models.
• ▹Authored and published peer-reviewed research in Springer Lecture Notes in Networks and Systems (LNNS), Scopus-indexed (ICTIS 2024).

Vishwakarma Institute of Technology

• ▹Developed an early-stage skin disease detection system using ML to assist in diagnosing conditions that are difficult to identify with the naked eye.
• ▹Addresses the shortage of dermatologists, especially in rural areas, by providing automated preliminary screening.
• ▹Uses image-based analysis for accurate, early, and objective diagnosis to overcome disadvantages of traditional visual examination.
• ▹Aims to reduce diagnostic costs and provide timely medical intervention for skin diseases.