GenAI @ Elastic | Machine Learning & Software Innovation

github

Adhish Thite  • Peerlist

The support team at Zuora needs to actively involve human agents for the resolution of support tickets. If the support ticket is related to a problem with Salesforce.com process, the support agent needs to request Salesforce.com org access to the customer. Requesting access and getting it requires a lot of back and forth thus adding to the ticket resolution time. If the time required to get the Salesforce org access is reduced, it will lead to a significant reduction in the overall problem resolution time.

In this project, we aim to use Machine Learning to solve this problem by using the concept of Correlated Topic Models, specifically the Pachinko Allocation Model.

Technologies - 

- Python 3.x
- Java v1.7
- MALLET/Spacy (NLP)
- Tensorflow
- Neo4J (Graph DB)

Automated Salesforce Org Access Checker

Proud to introduce ChatGPT Mini — a sleek command line interface (CLI) for ChatGPT that runs on your local machine. It’s the perfect solution for users who want to engage in conversations without breaking the bank or compromising privacy.

ChatGPT-Mini

A comparison of various optimizers on the CIFAR-10 dataset.

Optimizers-
- SGD
- Adam
- RMSprop
- Adadelta
- Adagrad

Demonstrated the accuracies of a model due to different optimizers, keeping all other parameters to default.

CNN Optimizers Comparison

Decoupled Neural Interfaces using Synthetic Gradients are greatly effective to improve the training speed in Deep Neural Networks.

Our research is to further improve these Synthetic Gradients by introducing a 'pre-training' module so as to observe the result of this pre-training on the actual training process.

Technologies-
- Python
- TensorFlow
- Keras

Improved Decoupled Neural Interfaces using Synthetic Gradien

Implemented a Deep Neural Network with KERAS in Python on the PIMA Indians Diabetes dataset. The model achieves around ~78% accuracy, which is in the Top-3 accuracies achieved by any ML/DL model for that problem.

Technologies & Frameworks used -
Jupyter Notebook (Python 3.x) on Google Colab
Keras

KAGGLE - DNN on PIMA Indians Diabetes Dataset

Developed a custom code in 'R' which uses 'multi-variate' or 'oblique' splits in each of the Decision Tree instance of a Random Forest structure.

The optimal split is  a 'linear combination' of two or more features selected at random in a single decision tree instance (oblique) and gives better space & time complexity than the traditional 'univariate' split.

Leveraged the standard implementation of the 'randomForest' package to make the oblique splits.

The output (error rate) is comparable to that of the custom R packages - 'oblique.RF', 'oblique.tree', 'PPForest (Project Pursuit Forest)'.

Technologies (Programming Languages) -

Python
R

Multivariate Splits in Random Forests (Forest-RC)

Built a Deep Learning model using (VGG-16 & LSTM) to generate captions for any image.

Technologies -
- Python 3.x
- Keras (VGG-16)
- NLTK (NLP & Bleu Score Calculation)

Future Goals - 
- Use a better model than VGG-16.
- Enable a REST API from a website to generate a caption for any input image provided by a user.

Neural Image Caption Generator

At Synapse, we employ modern artificial intelligence techniques based on computer vision¹ and deep learning to automate physical documentation verification. One of the applications is verifying government-issued identity documents, enabling us to identify and authenticate end-users.

PRODUCTION: ID Verification with Deep Learning

*** Real-time recognition of handwritten digits using the front camera of a laptop. ***

Implemented a Convolutional Neural Network model to train the MNIST dataset.

This model is then applied to images captured with a camera on the laptop, and then the handwritten digit is predicted.

Technologies/Libraries -
~ Keras (TensorFlow backend)
~ OpenCV
~ Dataset - MNIST (99.2% accuracy)

Real-Time Handwritten Digit Recognition (OpenCV & KERAS)

Digit Recognition from Sound

A neural network to correctly identify spoken digits. The model achieves 97% test accuracy.