Deep Learning | Soroco | IIT Delhi
6 Years of Experience
Bangalore Urban, Karnataka, India
+91 8**********
-
-
Not Available
I am a Data Scientist and I develop AI powered products. Currently, I am working on solving work graph optimisation problems using deep learning.
Soroco, SaaS/Cloud Product, Computer Software
BYJU'S
Axtria - Ingenious Insights
Soroco, BYJU'S, Axtria - Ingenious Insights
Job Title : Senior Machine Learning Engineer
Company name : Soroco
Period : November 2023 - Present
Location : Bengaluru, Karnataka, India
Job Title : Machine Learning Engineer
Company name : BYJU'S
Period : February 2021 - November 2023
Summary : Solving ed-tech problems using deep learning, natural language processing and computer vision .
1. Byju's QnA search
Built byjus QnA search using neural search
Having SSR of ~80% with ~2mn hits in a month. Generating ~$600k/month in revenue through sales leads.
2. Universal Search Enhancement
Developed universal search feature for Byjus app to fetch resources beyond just questions, utilizing the Bert Transformer for query intent recognition and entity extraction. Achieved an F1 score of 0.96, with Text SSR at 91% and Image SSR around 60% in production.
3. Hyper-Personalised Product Development
Developed hyper-personalised products using Large Language Models (LLMs), specifically GPT4, with prompt engineering, Langchain, and Pinecone for premium users. Products include Teacher GPT for problem-solving guidance, Math GPT for step-by-step math solutions, and ByjuWizz, a textbook-specific doubt solver.
4. LLM Evaluation and Fine-tuning
Evaluated and fine-tuned open-source Large Language Models (LLM) like LLaMA, LLaMA2, Falcon for applications in math reasoning and doubt-solving chatbots. The project involved fine-tuning LLMs using QLoRA and Peft framework. The LLaMA 13B model achieved an accuracy of 13% on GSM8k.
5. Finetuning Transformer
Fine-tuned the Bert transformer on ed-Tech datasets using siamese and highway network architectures. Implemented contrastive, cosine, and cross-entropy loss for similarity and classification tasks.
6. GAN based watermark removal
Built a cGAN-based model for watermark removal from solution images, aiding Byjus dB data ingestion. Utilized OpenCV, U-Net with pre-trained ResNet-101, PatchCNN
7. Upper Body Detector
Fine-tuned YOLOv5 for upper body object detection on annotated classroom data. This model is used for classroom engagement. Achieved mAP50-95 of 0.922 on YOLOv5l6.
Location : Bangalore Urban, Karnataka, India
Job Title : Data Scientist
Company name : Axtria - Ingenious Insights
Period : July 2018 - February 2021
Location : Gurgaon, India
Title : Introduction to Deep Learning
Period : August 2020 - Present
Summary : KEWRX4984FSE, coursera.org, http://coursera.org/verify/KEWRX4984FSE
Issuing Authority : Higher School of Economics
Title : Fundamentals of Scalable Data Science
Period : October 2019 - Present
Summary : QAUGBPJG3TLU, coursera.org, https://www.coursera.org/account/accomplishments/verify/QAUGBPJG3TLU
Issuing Authority : IBM
Title : Embedded & Robotics Basics
Period : June 2014 - Present
Summary : HP/STP/2014/NT2819
Issuing Authority : HP India Sales Pvt Ltd
English (Professional Working), Hindi (Full Professional)
Show More