Ashutosh Tripathi

An NLP engine that performs various tasks, including sentiment analysis, text classification, risk analysis, and real-time dashboards. I built this artificial intelligence model using Python 3 and TensorFlow. I was the project owner and worked on the project alone.

The model is based on a deep recurrent neural network (bidirectional LSTM) and attention mechanism. The model performs three tasks. First, it assigns a sentiment characterization to the text, which can be negative, neutral, or positive. Second, It assigns a predefined category to the text (I cannot mention the categories due to privacy agreement). Third, it assigns a risk associated, i.e., whether it's low, medium, or high.

The model is then transformed into a REST API using Django and Flask API. Also, the model is connected to a PostgreSQL database and dashboard. The app is then wrapped inside a containerized Docker solution which is then placed in a Kubernetes cluster.

A time series forecasting engine for predicting the capacity of telecom radio cells. Future capacity is vital because if the telecom radio cells load increases, there's a corresponding degradation in the user experience like slow data rates, call drops, and so on.

The goal was to better prepare for the future. I was the project owner and worked on the project alone.

The model is combination of regression model and time series model. KPIs like data traffic and number of users are forecasted using combination of models like ARIMA, Holt-Winters, Facebook Prophet, and deep learning-based models. This is done for each individual cells.

We also built a regression model using recent data for predicting KPIs like PRB utilization and user throughput against KPIs used for forecasting. Using combination of these two forecasting and regression models, we calculated capacity. The model is then converted into an application using the REST API, Docker, and Kubernetes.

A sentiment analysis engine for predicting positive, negative, and neutral sentiments on book review data. I was the project owner and worked alone on the project.

I built two models. The first dealt with reviews that were less than 250 words. The second was for reviews greater than 250 words. The first model was based on the XLNET transformer. A Hugging Face transformer library was used to build and fine-tune the model. The model gave results with an accuracy of 91%.

The second model was built using Bi-LSTM and with TensorFlow and Keras. The review was first tokenized using sentences and fed to LSTM as a time series data point. I then used Universal Sentence Encoder to get pre-trained sentence embedding. This model gave results with an accuracy of 90%.

A natural language machine translation app designed to convert Hindi to English and vise versa. This was my end year project during college. This is based on sequence-to-sequence architecture.

It uses an encoder and decoder. The encoder and decoder are based on bi-directional LSTM. It also uses a global attention mechanism to better learn long-range dependencies and focus on more relevant inputs from sentences. The model was built using TensorFlow and Keras.

This is a biological cell image classification project. Low-grade squamous intraepithelial lesion (LSIL) is a common abnormal result on a pap test. It's also known as mild dysplasia. LSIL means that your cervical cells show mild abnormalities.

The goal is to detect whether the cell is LSIL or normal. It's a binary classification task on image data. There were, in total, 10,000 labeled data which 50% of data was LSIL and 50% was normal. The model was built using Keras and TensorFlow ad a VGG 16 model was used for fine-tuning. The overall accuracy was 97%.