Amitesh Sharma

JPMorganChase, Senior Data Scientist

Working on large language models (LLMs), generative AI, and agentic frameworks to push the boundaries of AI capabilities.

Fidelity Investments, Data Scientist

• Survey Summarization & RAG: Developed an LLM-based summarization tool for user survey responses, which generates a summary and identifies top themes and top negative items. Developed a retrieval-augmented generation framework to gather deeper and specific insights for the same. Also developed a POC to create Knowledge graph-based RAG(retrieval augmented generation) system to improve retrieval time.
• App Placement using Recommender Systems: Experimented and deployed Multi-armed bandit-based recommender systems to order components and contents(video collections on Discover page) with a solution to handle the cold start problem.
• Growers and Reducers: Created a modelling ensemble to identify customers that were more likely to increase their assets(growers) and likely to take out their assets(reducers) in the next year by analyzing profile and interaction data(clicks, calls, etc) so that business can intercept at the right time and target the right customers.
• Explainable AI Library: Created an in-house library used by data scientists across AICOE that generates HTML reports(which includes performance reports, explainability, drift, and counterfactual analysis) that support classification(binary and multi-class) and regression use-cases for Torch, TensorFlow, and scikit-learn models.
• Lead Generation Framework: Developed a generic Deep learning framework that generates leads as per the given KPI criteria and rank orders the leads based on customizable metrics users provide. Created the featurization layer by creating time interval weighted embeddings followed by a modeling layer that takes sequential features as input.

Gaian Solutions, Data Scientist

• Question Generation: Developed a system to generate MCQ questions, True or False questions, and Descriptive questions (models based on T5 and GPT2) from a given text along with an automated validation model using a QA model(T5-based) to answer the generated questions to filter out unsatisfactory questions.
• Video Summarization: Developed a model for video summarization(in text) to find context for ad placement. Bi-modal Transformer is used for Multi-modal Dense Video Captioning by applying the model in a bidirectional manner through the video.
• Break Yield Optimization: Employed probability-based constraint formulation and reinforcement learning to generate optimal custom weekly advertisement schedules for maximum profit for the broadcaster while ensuring that the contract does not fail.

Quantela Inc, Data Science Intern

• Multivariate Time Series Imputation & Forecasting: Compared GAIN - Missing Data Imputation using Generative Adversarial Nets and Multi-directional Recurrent Neural Networks for missing sensor data.
• Univariate Forecasting: Developed end-to-end solutions for time series forecasting using Neural Beats.

M.Tech in Machine Learning and Computing

Indian Institute of Space Science and Technology | GPA: 8.63/10

B.E. in Computer Engineering

Rajiv Gandhi Institute of Technology – Mumbai University | GPA: 7.78/10

Adaptive Traffic Signal Control

Created a policy structure by network sharing for different topological road structures to reduce training time and obtain signal phases for large maps. Leveraged multi-headed attention for inter-intersection cooperation and used a deep Q network for selecting traffic signal actions. (M.Tech Thesis Project)

Comparative Analysis of LSTM Variants for POS Tagging

Implemented a POS tagger using LSTM and its variants, and conducted a comparative analysis to determine the best performing architecture for part-of-speech tagging. (Mini Project)

Learn More

Image Segmentation using K-Means, Fuzzy C-means and Mean Shift Algorithm

Various techniques for image segmentation were compared with each other using the ground truth.

Learn More

Real Time Object Detection, Recognition & Price Estimation

Developed a real-time system that captures webcam frames to detect and recognize objects using a pre-trained model. The system predicts the price range for each identified object and provides purchase links via vendor APIs. (B.E. Major Project)

Paper: Detection and Recognition of Objects and Providing Purchase links using APIs

nbeats-forecast

nbeats_forecast is an end to end library for univariate time series forecasting using N-BEATS (Published as conference paper in ICLR). This library uses nbeats-pytorch as base and simplifies the task of forecasting using N-BEATS by providing an interface similar to scikit-learn and Keras.

Git Link: https://github.com/amitesh863/nbeats_forecast

PyPI Link: https://pypi.org/project/nbeats-forecast/

Pune Urban Data Exchange (PUDX) Datathon

Member of the winning team for Pollution Exposure Data Analytics. Air Quality data, Rainfall data and location of the sensors were given by the Robert Bosch Centre. We performed Seasonally Decomposed Missing Value Imputation using Kalman filter for missing time-series data and trained Neural Beats model for univariate forecasting of PM 2.5 values.

Hackathon Link

Certificate: Verify Certificate

Programming Languages

Python, MATLAB, C, C++, SQL

Libraries & Tools

PyTorch, TensorFlow, NumPy, Pandas, SciPy, Matplotlib, Scikit-learn, OpenCV

Deep Learning - Part 1

NPTEL – Verify Certificate

Machine Learning

Stanford University – Coursera – Verify Certificate