Feb 2024 – Nov 2025 · 1 yr 9 mos · Gurugram, Haryana, India · On-site

• User-Ticket-Service
• Developed hierarchical AI Agents to automate user ticketing service using Google-ADK and LangGraph.
• Created MCP servers with FastMCP that provide context to the multi agent systems using our legacy APIs and interact with remote JIRA and Salesforce MCP tools to orchestrate ticket resolution.
• Architectured query gateway and history management services to accurately route and manage query state across sessions.
• Vaarta-Vedh
• Developed a financial advisory chatbot using NLP and Generative AI that provides personalized tax and financial advice based on users earning and spending patterns and can handle multi-turn and multi-topic conversations.
• Developed text-to-NoSQL query generation module using LLMs, which processes 200+ unique categories/subcategories, payment methods, merchants and associated banks to create accurate NoSQL queries.
• Designed a RAG based caching system to store and query frequently asked user questions. Performed template matching with hybrid search for higher accuracy, and reduced latency by 70-80% and costs by 90%.
• Created a response generation module that uses a Llama3.2 3b model fine tuned on custom SOPs and guardrails to create structured responses. Trained with PEFT(QLoRA) and served using vLLM.
• Built an automated evaluation tool, lens-test-suite, to check for drifts and evaluate system responses.
• Bank-Statement-Parser
• Developed a scalable credit card statement parsing service that uses OCR and large language models
• (Azure OpenAI) to extract and parse text data and send bill payment notifications to users.
• Created FastAPI service to consumes requests in batches from Kafka queues and processes them asynchronously using Asyncio.

Sep 2022 – Feb 2024 · 1 yr 5 mos · Buffalo, New York, United States

• Worked on multimedia news forensics under Dr. David Doermann. My research involved detecting tampered and inconsistent news articles.
• Face-Morphing
• Developed an image morphing system that uses latent diffusion models (Control-Net) to generate high quality face morphs of interpolated best pair images.
• Implemented latent interpolation and CLIP-based scoring to optimize image generation quality.
• Manipulation-Detection
• Developed manipulation detection system to detect, localize, and label tampered news articles.
• Trained a CNN model using differential images with ResNet-50 as backbone, to detect compression artifacts with 93% accuracy.
• Utilized NEDB-Net to extract noise and edge-based features to localize manipulations for the tampered regions.
• Fine-tuned a custom Yolo-v8 model to detect objects in the localized regions and label them into 18 categories.
• Engineered text-transformer tool for controlled entity and parts of speech replacements using SpaCy and transformer (BERT) models, enhancing data preprocessing for downstream AI tasks. Used co reference resolution to maintain context in long texts.
• Multimodal Inconsistency Detection
• Developed a Python tool for controlled entity replacements in long texts using SpaCy and NLTK.
• Conducted data cleaning, preprocessing, and exploratory analysis on news datasets.
• Detected inconsistencies in multimodal news articles with entity mismatch and contradiction detection algorithms.
• Built a retrieval-based chatbot on a self-curated Reddit dataset, leveraging Solr for text indexing and searching.
• Implemented embedding based query matching for ranking and co-reference resolution to maintain short context.

May 2022 – Aug 2022 · 3 mos

• CSE 702/701 : Automated Analysis of Sporting Event Videos
• Instructor : Prof. David Doermann & Prof. Nalini Ratha.
• Guided students to develop deep learning projects for video analysis in sports domain for CSE 701/702.
• Conducted classes, graded assignments, and reviewed students’ technical presentations.
• Independent Research
• Point cloud segmentation for autonomous driving using local feature aggregation on PandaSet.

Feb 2018 – Aug 2020 · 2 yrs 6 mos

• Trained a Random Forest model to classify independent and dependent system components with 97% accuracy.
• Built a part price prediction module using linear regression for refrigeration units, supporting faster data-driven quotations.
• Managed data workflows and CI/CD pipelines using Jenkins, Git, Airflow, and PySpark.
• Developed employee task delegation system to manage production workflow data using Python and React.
• Built REST API to parse delegation data from JSON files received in hourly batches with Flask and Python.
• Designed scripts to extract, parse and transfer millions of parts data across 6 pricing interfaces using PLSQL.
• Reduced batch transfer failure rate in batch jobs by 20% through RPA automation and monitoring.