Machine Learning

Structuring Machine Learning Projects

Neural Networks and Deep Learning

Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization

Convolutional Neural Networks

Public Relations Event Coordinator

Public Relations Intern

Sponsorship Coordinator

Convenor

Operations Coordinator

Student

Bachelor's degree

Under Prof. Arun K. Pujari Vice Chancellor CURaj, I worked on the 3D reconstruction problem using deep learning. 
Different works are Visual Hulls(3D reconstruction algorithms),3D representations like voxels,octrees,Object Pose estimation(6 DOF),Semantic Segmentation.
Constructed the pipeline of complex networks that include the combination of different CNN’s, Visual Hull construction algorithms. Different Networks are tested and modified.
Trained CNN’s model in a GPU with the environment.
Used Py-Torch framework for implementing the model.Other libraries were numpy,matplotlib,PIL,OpenCV,Matlab.
Different pooling sizes are tested as 7*7,3*3, 1*1 with skip connections and batch normalisation to different units. Finally, a basic understanding is drawn to put the different techniques like the batch norm, skips to the CNN units.
The pinhole camera model is used. The 3D object is reconstructed in the unit-length cube and focal length is assumed.

Research Intern

Designed a web-based online portal where one can study coding and gives test.  My work includes development using JAVA. Web development and desktop Application development include my Projects. JAVA Backend is based on Servlets,JSP,EJB,JDBC.

Intern

Worked on Text Recognition in OCR engine.
Used Service-Stremmer for increasing different model inference API’s by 10x. Deployed Face Det,Face Match, Liveliness, Gesture etc. 
Designed a neural based technique to filter out 5000 dissimilar images out of 50000 images.
Other work includes Arabic Text Recognition and its data preparation using the same pipeline.
Designed a pipeline that used  Signzy  OCR system and then upload the results to nebula platform for data preparation so that Human annatotars can annotate them and Signzy OCR can learn from its own mistake.
Worked on Blink NN have CNN+LSTM’s in Time series fashion for analysis of each frame.
Worked on Face Detection and Face Match DL models.

Computer Vision And Deep Learning Intern

Holds ownership for various products like all Segmentation use cases, Video, and facial products like face deduplication, face match,etc. Experience in designing AI-based Products, client-based projects,microservices,etc.

Core AI SDE-2

Working on digital image tempering i.e copy paste forgery. Build and handle face detection and recognition engines(1:1 match),Optical character recognition(OCR)(CRNN models design) ,  deep-face-authentication(1:N match),aadhaar and  ekyc forms 8,12 digit masking(both OCR and segmentation methods),  cheque extraction(NER).
Work on the Pixelation effect and text analysis in Image quality analysis and Screenshot detection/classification. Designed many end to end ai micro services and products. Handy with docker, flask and redis and different servers(aws,e2e, azure ) for deployment.

Core AI SDE

Jagriti Yatra is a 15-day long, national train journey that takes you 8000 kilometres across the length and breadth of India, to understand and build the India of smaller towns and villages through enterprise.

 Yatri

Research Paper Got selected at Coling22. Long paper, Oral presentation. 
Summary :
Our work on Context-sensitive Compound Type Identification got accepted in Coling'22 for Oral Presentation. We propose a novel multi-task learning architecture which incorporates the contextual information and uses morphological tagging and dependency parsing as two auxiliary tasks. Experiments on the benchmark datasets for Sanskrit show 6.1 points (Accuracy) and 7.7 points (F1-score) absolute gain compared to the state-of-the-art system. Further, our multi-lingual experiments demonstrate the efficacy of the proposed architecture in English and Marathi languages.

Research contract

Project Link: https://omdena.com/projects/ai-wildfires/
Blog: https://www.omdena.com/blog/junior-data-scientist
We are working for social good for the detection of fire outbreaks in the Brazilian forest. The dataset is provided by Sintecsys.

Junior ML Engineer

● Led 7 Different Products. Build 5 highly impactful products from Scratch. 
● Product: Led end-to-end on the device(Web, iOS, Android) face photo
capture with live feedback. Compressed a model to 1.5MB to use on the
web. This product reduces the visa reduction rate by 40%. Carry out various experiments to monitor user drop-offs.  
● Product: Led the product, design and engineering implementation of the
review service for Atlys build trust and conversion. Expending this to i18n,
Indian, forex, real reviews fetching from Play Store, app store, refining customer feedbacks offline. Brings 25%
growth in user conversion improving trust.
● Product: Designed AI Chatbot that answers all your visa-related queries.
And build AI Conversation agent for customer support, reducing Customer
support load from 60% to 5% support tickets and human involvement.
● Build AI Searches
●Design a chat-based document collection bot to collect documents from user.
● Expending it to remove human intervention.

ML & Product

1. Product1: Developed a virtual weighing machine using simple mobile camera videos. Built a ResNet+tabular combined model achieving an MAE of 114gm.
2. Models: Tested out various video-based models and transformers such as VIT, Vanilla,DAFT,etc combining tabular information for multitask predictions.
3. Optimization: Reduced model size by 92% and increased speed by 91% for low-end devices using pruning and quantization in pytorch.
4. Accuracy Improvement: Integrated visual and tabular data for superior predictions, improving accuracy by 20%.
5. Model Analysis: Explained AI behaviour through attention and weight distribution analysis.
Data Quality Initiative: Reduced data collection wastage by 30% through video quality checks.
6. Product2- Pregnancy Report Stratification: Developed an end-to-end tabular solution achieving 80% specificity and 70% sensitivity. Managed overall development of the product. 
7. Developed a Women's Consultancy Chatbot Using Various LLM Models and LangChain wh

Associate ML Scientist 

Under Advance Geometric Computing Lab I am researching on 3D Geometry. Moreover, I am researching to develop 3D object reconstruction techniques using AI.

Co-Founder

I have co-founded Hyperint. We have 2 products live. Selfie Validator and WordOfMouth. Working on one more product.

Know about Ashish Gupta's work experience, education, and much more. Ashish Gupta