Sep 2022 – Present · 3 yrs 9 mos

• Built various Multi-Agent AI applications in the customer support chat space, using GCP Agent Development Kit, to boost containment by 55%, and service turnaround time by 30%.
• Introduced a novel LLM Confidence Measure mechanism using LLM API internal metrics to improve Generation Trustworthiness by 75%. A research publication is currently under the works in our research division of the company for this innovation.
• Developed an improved RAG System by implementing rejection mechanisms at multiple points in the standard pipeline to improve accuracy and reduce hallucinations by 25%.
• Finetuned custom lightweight language models and deployed them on GCP Model Registry as an endpoint, replacing vendor LLM intent classification and reducing costs by 90%.
• Developed multiple Prompt Engineering-based GenAI use cases, with Large Language Models (LLM) such as ChatGPT, Anthropic, Cohere, etc. tailoring solutions to diverse customer and developer personas.

Jun 2022 – Aug 2022 · 2 mos

• • Overlap analysis on textual data for a multiclass classification problem. Experiment on various preprocessing techniques to minimize class interaction and hence boost the accuracy of an AutoML provision from SAP (Data Attribute Recommendation).

Nov 2021 – May 2022 · 6 mos

• Exploratory Data Analysis phase on F5 Load Balancer Telemetry data with HANA Infra Analytics.
• Conclusive feasibility check for an NL2SQL project. Assessing various pretrained models like GPT-J, T5 + WikiSQL, SQLNet, etc. on a limited training dataset.

Sep 2021 – Oct 2021 · 1 mo

• As part of the Tech-Startup Hackathon event InnVent, built a Text-to-Command utility using a Custom Transformer Sequence-to-Sequence architecture written with Keras (Tensorflow) that was trained with a hand-curated Dataset of Natural Language text and corresponding commands that were specific to cloud providers (AWS and GCP). Attained ~80% accuracy, especially given the limited training data for the Minimum Viable Product.

Jun 2021 – Dec 2021 · 6 mos

• Worked with the SAP API Services team to develop a Semi-Supervised Anomaly Detection service for the API traffic that is received across various Tenants. Multiple approaches were tested for this; I majorly worked on developing a Deep Learning model that uses an LSTM Autoencoder architecture, on a reconstruction error based approach. This used Tensorflow (Keras). Attained ~90% accuracy over 1 year of training data.
• Helped analyzing the performance, and hence finalizing the usage of other decomposition based libraries like Facebook-Prophet and Microsoft's SR-CNN. All three models are Dockerized containers carrying a RESTful Flask application each, and deployed as separate microservices with an additional service orchestrating the detection of outliers in incoming traffic.

Apr 2021 – May 2021 · 1 mo

• Worked with the mobile services team as an ML Engineer on the month-long hackathon Innovision to create an NLP Multilabel Classification pipeline. This project used Bidirectional LSTM using Keras (Tensorflow) in order to identify the type of service required based on a given sentence and predict a class.
• Deployed above application as a Flask service on CF. The service had additional endpoints that could allow re-training the model with more or completely new data. The aim of this project was to showcase to customers all the offerings that the MDK Landscape can provide.

Aug 2020 – Mar 2021 · 7 mos

• Implemented a feature enhancement that significantly minimized pipeline freezes, from 90% of the time to <10% of the time, which were caused by dataload limitations involving sizes of 1m+ of records. (SpringBoot, Python)
• Took over various product bugfixes and delivered them in record time. (SpringBoot, Angular, Python)
• Tested various Time Series forecasting methods that SAP HANA Db PAL (Predictive Analysis Library) had to offer, wrote a custom Python pipeline that analyzes about eight of said algorithms and asses the performance of each across the entire dataset containing various customers, commodities and their expenses.
• Post analysis of best algorithms, implemented an expense forecast Proof of Concept with custom stored procedures written on HANA invoked via a SpringBoot backend.