Experimented with RL Algorithms like variants of DQN, Vanilla Policy Gradient, Proximal

PG, TRPO in OpenAI Gym environment for both Discrete and Continuous action space on

toy Car for steering its wheel.

Developed Computer Vision based Real time counting software for forging plants using IP/Analog Camera and deployed it on the Raspberry-pi. Implemented and Trained both Custom and pretrained SSD based model for object detection and modified TensorFlow object detection API source code and trained SSD based model, later successfully build quantized model using TFlite.

Implemented Object Detection and Tracking model that can run on Edge devices like Raspberry Pi at 10FPS & alerts to enable immediate actions. The light weight Object detection model has size nearly 700kB implemented based on SSD paper . The Object Tracking model implemented based on Deep-SORT paper. All these generated events sent to Zoho Analytics Cloud for data visualizations.

Improved the safety, security and efficiency of the plant. Red Zone indicates the Danger Zone in the Videos, if person detected in that zone save the 30sec clip and shows it on the Dashboard. Also check for the Helmet like whether person is wearing Helmet or not and alert accordingly.

Every ATM has 2-3 camera installed, and they record video all the time whether person inside ATM or not, and the data can be more than 2-3GB and analyzing these video is hard to do manually, if bank try to look for some event. So just record when person inside ATM (motion detection + person_detection). Also look for other events like Multiple People inside ATM, Person Wearing Helmet, Person face is not visible etc and send the alert right away. Also send the binary signal alert using Relay board Connected to RPi, and this Relay output goes to ATM sensor board (fire, smoke, vibration, thermal etc.) which sends the signal to the Central ATM Server for quicker action in case of Video loss/or someone broke the camera (Major security concern). The 2 top cameras are synchronized and one of them checking for first motion, and if it is true then start the Person detection model and light-weight Kalman filter tracking. The front camera detecting either face or helmet.

Store only the relevant data. Reduce the cost of manual monitoring of ATM. Alert any illegal activity inside ATM in Real-time. Both of the Model runs at 5FPS with max 90% CPU utilization.

The goal was to detect presence of an object by analyzing the sudden change in value of ALS Sensor data. There is no analytical method to determine the dip, it could be from 700 to 300 or 1200 to 700 or 1200 to 300. So by using Time-Series Prediction model like LSTM which would tell us the presence Anomaly is where reconstruction error is large. And by using as much as 4 ALS Sensor and method of Triangulation we can approximately locate the presence of an object.

By using existing sensors, I was able to approximately locate the presence of object. Either we have to use Camera or high-end Sensors like LIDAR. This saves the cost by not changing the existing structure where this measurement has to be taken.

This website that let New User to make Transaction on Bitcoin Testnet. It
makes new user Understand how bitcoin transaction works under the hood. Currently it supports Pay-to-PubKey Hash (P2PKH) transaction format. The Bitcoin P2PKH transaction protocol written from scratch in JavaScript. The Crypto algorithm Elliptic Curve, SHA256 Hashing algorithm, RIPEMD160 all of them are written from scratch in JavaScript.
The website is written in Plain HTML and CSS.

Bored Apes images are generated using GANs. The model is trained on 10000 images taken from BAYC Opensea website. The Smart contract is
written in Solidity deployed at Rinkeby testnet. The website is written in React and I've used various Node.js packages. I've used eth-brownie for smart contract deployment and calling it's function.

Successfully deployed Solidity Smart Contract. Currently minted 20 Apes. It also supports MetaMask wallet.