To make the reactive cctv cameras proactive using a event driven system which runs the Ml algorithms on edge device and raise events.

Tech used : AWS cloud ec2, ecs ,lambda, cloudfront, cognito,Dynamodb, RDS kinesis video stream, iot Core, grpc, tensorRT, YOLO

• Developed an end-to-end system encompassing a user-accessible web application built with Next.js, a cloud infrastructure hosted on AWS, backend APIs implemented in Python, and edge software utilizing GRPC modules.
• Designed and implemented the cloud infrastructure on AWS, utilizing services such as EC2, S3, and RDS to ensure scalability, security, and reliability
• Developed robust and efficient backend APIs using Python, enabling seamless data flow between the web application and the cloud components.
• Ensured end-to-end security and data privacy by implementing encryption protocols and access controls across the system.

To build a dashboard to show the realtime data produced by multiple IOT sensors and apply a anomaly detection alogritm on the data to detect and show anomalies in time series data.

Tech used : AWS EC2, Kinesis datahose, data stream, Athena DB on S3, Grafana for dashboarding, lambda, Autoencoders

• Developed and deployed an efficient data pipeline to capture real-time data from IoT sensors and process it for analysis.
• Implemented a seamless integration between IoT devices and the pipeline, ensuring reliable and continuous data ingestion.
• Designed a scalable architecture to store the raw sensor data in Amazon S3, enabling cost-effective and durable storage.
• Leveraged AWS Athena to perform ad-hoc queries directly on the raw data stored in S3, providing real-time insights without the need for preprocessing or data transformation.
• Configured dynamic dashboards in Grafana to monitor and analyze key metrics and trends derived from the IoT sensor data.

To build a recommendation system based on the interests and entities the user interacted with.

Tech used : Recommendation system, NLP, DOC2Vec, Word2vec

• Developed and implemented an advanced article recommendation system within jio health hub app, leveraging user interactions with various entities such as diseases, medicines,symptoms …
• Implemented content filitering techniques to solve the cold start problem.
• Implemented collaborative filtering techniques to analyze user behavior and preferences, identifying patterns and similarities to generate accurate recommendations.
• Integrated natural language processing (NLP) techniques to extract key information from articles and user interactions, enhancing the relevance and quality of recommendations.
• Implemented algorithms and strategies to ensure diversity in recommendations, promoting exploration and discovery of new health-related topics.
• Utilized feedback loops and metrics to evaluate the performance of the recommendation system, fine-tuning algorithms and parameters for optimal results.

To build a an api that can extract medical entities from free text and map them the appropriates entites in the DB.

Tech used : elastic search, Fastapi, NER

• Developed, deployed and maintained a Fastapi on top of elastic search which does NER (Named Entity Recognition).
• This system is used across the app anywhere there is an entity resolution required.
• This API also support autosuggest, look ahead search functionalities
• Built a POC for scalable unsupervised learning based named entity recognition for semantic search usecase.

To solve master data management problem with Machine learning in different stages. Key challenge is to do Item duplicate detection, Category (noun modifier) Identification, Attribute extraction on item description (short text)

Tech used : Elastic search, mongo db, Azure Web-app, Azure storage tables, Azure Data Factory, Support Vector Machine, Named Entity Recognition, POC on LUIS

• Built Azure data factory pipelines run batch job on millions of rows.
• Built feedback loop for extraction alogo to updated extraction regex based on user decisions
• Started with simple SVM classifier for Category detection and upgraded to deep learning based(embeddings training) semi-supervised identifier
• Trained a custom NER attribute extracter
• Improved master data search to give better contextual results by identifying parts of search query
• Designed, built CICD pipelines for ease development and deployment.
• As a part of supplier details validation, implemented all rules to validate the details using python. Also implemeted realtime address validation using google API

To automate extraction of metadata, clauses from scanned, signed contracts. which helped in reducing the manual efforts of internal team by more than 30 %. Which was then productized so it can be used directly by clients .

Tech used : py-tesseract, Azure web-app for containers, Azure DevOps pipelines, Azure cognitive services, Keras, Docker, Flask, CNN, RNN, Decision trees, Mongodb (Azure Cosmos DB), Azure tables services, Azure blob storage.

• Trained Image classification model using CNN to identify different types of pages in contracts.
• Trained Decision tree model to decide the type of contract from the first 2 paragraphs of the contract.
• Developed and trained a Recurrent neural network model(LSTM) for capturing address from natural text.
• Developed and trained a multi-class sentence classifier for identifying types of clauses in a contract.
• Built, Deployed and maintained APIs to extract metadata using all these machine learning models.
• Designed, built CICD pipelines for ease development and deployment, Designed and built an automated feedback mechanism to train and deploy models automatically.
• Engineered Tables extraction and structure maintenance from scanned documents after OCR.

Given an image, Detect known objects in the image and find similar images in the database of images.

Tech used : Deep Learning using Keras, Transfer learning, CNN, Object-detection and Image Classification, KNN for Finding closest images

• Used pre-trained Res-net 50 architecture to encode the images, Used Modified K-nearest neighbor to find the images for the recommendation.
• Used classification models to decrease false-positive rate and object detection (Yolo) to identify multiple objects from one image and recommend accordingly.
• Deployed the entire project with Models as an API and made it as a framework to update the database.

Given Sale or revenue of previous years of a US-based retailer forecast the sales for next year.

Tech used : Implemented Linear, Polynomial regression, LSTM (RNN) and Decision tree regression with combining Macroeconomic, weather data along with the given data. Preformed grid search for hyperparameter tuning.

• Ranked 2/93, Won Gold for getting the lowest RMS error

Designed, Built and Deployed multiple POC’s, working prototypes in very short time to demo for clients

Tech used : YOLO - CNN, RNN, chatbots

• Setup a matter most server on an AWS EC2 instance, deployed 3 different chatbots, which can be triggered in a group chat
• Trained an RNN based question and non-question classifier using the SQUAD dataset which is used to automate the questions extraction from a chat transcript.
• Built a real-time dashboard using dash plotly which shows the graphs and counts of people and vehicles.
• Built a YOLO based object detection model to analyze the CCTV footage at a gate and count people, vehicles going IN-OUT

Designed, built and tested the prototype of a laser range finder for detecting breakthroughs in underwater boreholes (Subsea level).

Tech used : Used CMU camera for object detection, Arduino microcontroller, Matlab for real-time 3D visualization, and designed a Solid works CAD model.

• Master’s Project

we designed and made a line follower robot with an option to various control parameters in PID system.

Tech used : Worked on Arduino Microcontroller, Tuned PID control parameters.

• Used to demonstrate the performace of a system with varying P,I,D values
• Master’s sem Project

Built a Biped robot using servo motors and Arduino.

Tech used : Solidworks, Arduino

• Programmed and used Arduino board to control servo motors, used Solidworks application for designing the links, performed structural analysis of Biped

we designed, engineered and assembled the working model of a machine, which replicates the three-hand motions through which RUBIK’S cubes are solved.

Tech used : Expert with Lathe, Milling, CNC, Cutting and Drilling Machines tools.

• This project was awarded the best project in the class by IIT Kanpur.

CS231n, CS224n, Advanced Engineering Mathematics, Discrete Mathematics, Robotics, Modern Control Systems, Linear Algebra, Probability, Mechatronics, Dynamic systems, Inverse kinematics.

Rock-paper-scissors (Ultrasonic sensor, LCD screen, Servo motors), Go for green(Used Ir sensors and LEDs), obstacle avoiding robot (Ultrasonic), Step lights(pressure sensors, LED’s), Robotic Arm (3 Servo motors).