Seeking a challenging position where I can apply my knowledge, experience to work towards achieving the organizational goals with continuous mutual improvements.
“STP For Wires Processing and Internal Transfers” is a change to how the wires and internal transfers are processed in the MUFG bank. This change involved making several enhancements to the exisiting applications to increase UX and also removing several legacy applications. New applications were added to decrease the TAT and help process the wires faster.
Klik solution helps financial institutions and major corporations simplify receivables management. It provides private label cash management products and end-to-end outsourced solutions allowing companies to better serve their clients - while minimizing cost and complexity. It helps companies grow, increase market reach and streamline costs with our hosted, cutting-edge receivables management solutions.
70.01%
74%
77%
Built a ResNET50 from scratch using only tensorflow and numpy. The model was then trained on a small but complex dataset of only 120 images.
Deep neural networks can often learn complex features and representations. In our cat vs non-cat dataset which was very complex but was of very small size (only 120) images gave us great results using ResNET50 giving us an accuracy of nearly 70 percent without any optimizations. The model used the tensorboard and model checkpoint callbacks to monitor the validation accuracy.
At first, the model which I used was the CNN model which was doing great on the training dataset but was not performing well on the test dataset which clearly was the case of overfitting. The problem was the small dataset of only 120 images. I tried using data augmentation but that didn't help in great advancements. Then using the concept of transfer learning and using the VGG-16 model I was able to attain almost 90 percent accuracy and correctly classify the cats.
This was the first project which was using the CNN model. The model was debugged using grid search like process. The model was then further tuned using tensorboard which helped greatly in visualizing all the metrics associated with the model. The dataset was also pretty huge (24946 Images) so it also helped me in working with a large dataset
The home automation project using Internet of Things was our final year project. The project enabled the user to control various appliances in his/her house through the use of mobile app. The project required a constant internet connection and the devices could be controlled remotely.
The devices were connected through arduino board which was in turn connected to the internet through an ethernet shield. The arduino was communicating to an web server created by us. When the user pressed "Switch On"cbutton on the android application the command was passed to the web server and thus the applicance was switched on.
This was my first Django project which converted a long URL of any length to a tiny URL by using decimal to base 62 conversions. The stats are stored in the database and the site is stylized using bootstrap.
The Django URL Shortener as the name suggest is a URL shortening service which was completely built from scratch by me. It uses a base 62 to decimal conversion to convert a long URL to a short one.
Built a neural network from scratch using only numpy. The neural network was of multilayer perceptron which was continuously enhanced using Adam optimizer, momentum, rms prop, and Xavier initialization. The model attained an accuracy of 98 percent.
Starting from scratch was a very good experience as I got to see my model continuously improving with each enhancement. With just a little bit of tinkering, I was able to increase the accuracy from 95 percent to 98 percent. As the whole model was built from scratch I was able to get into the nitty-gritty of every topic.
Using the webcam of the laptop the model was able to detect fidget spinner with a bounding box around it. The dataset was of 300 images and was scraped from google images and was labeled by me. The images and its respective labels were then used to train the tiny YOLO model.
The project started with collecting the dataset which was images of the fidget spinner scraped from google images. The images were then labeled drawing bounding box around each fidget spinner in the image. Then the model was trained using only datasets of 300 images. When the model was run on live webcam feed it was successfully able to identify the fidget spinner without any difficulty.