At Pulp Internet, an early-stage AI startup, I take on a multi-faceted role spanning full-stack development, AI research, and data engineering. Working in a fast-paced startup environment has given me the opportunity to lead research initiatives while building and optimizing AI-driven products. I have developed AI-powered chatbots, metadata enrichment pipelines, and scalable knowledge retrieval systems, integrating technologies like OpenAI, Pinecone, Neo4j, and AWS to improve efficiency and user experience. Additionally, I have designed GraphQL and RESTful APIs, implemented role-based authentication, and optimized Next.js-based frontends to enhance the overall product ecosystem.

Beyond full-stack engineering, I have contributed to data pipeline optimization and intelligent search retrieval, engineering LLM-driven recommendation systems, RAG-based search, and AI-powered media processing. My work on knowledge graphs and metadata enrichment has increased retrieval efficiency by over 60%, while my serverless cloud solutions have significantly reduced API latency. At Pulp, I have not only deepened my expertise in AI and scalable architectures but have also taken end-to-end ownership of product development, from research to deployment, shaping how AI-driven applications can enhance user interaction and business intelligence.

Under the guidance of Professor Banerjee, I am actively researching multi-agent argumentation in AI systems, focusing on how LLMs engage in structured debates and reasoning tasks. Our work explores systematic vs. heuristic reasoning in AI, analyzing over 10,000 argumentation sequences across 12+ large language models (LLMs). By designing multi-agent debate simulations, we aim to understand how AI systems construct, challenge, and defend arguments in adversarial and cooperative settings. Through controlled prompting experiments with 200+ structured tests, we have observed significant improvements in systematic reasoning and logical coherence, refining our approach to AI-driven deliberation models.

A key aspect of our research is identifying the limitations of current LLMs in logical argumentation and developing methods to enhance reasoning depth, fact-checking capabilities, and adversarial robustness. We are investigating prompt engineering techniques, structured reasoning frameworks, and token efficiency optimizations to improve AI-generated argument quality. Additionally, we are exploring how knowledge graphs (Neo4j) and retrieval-augmented generation (RAG) techniques can enhance AI-driven argument retrieval and support claims with factual evidence. This ongoing research aims to contribute to more transparent, explainable, and logically rigorous AI systems, bridging the gap between human-like reasoning and computational argumentation.

At WebGen, a Vertically Integrated Project at Stony Brook University, I contributed to the development of genomics data pipelines and real-time visualization tools for biomedical research. My primary focus was on processing and optimizing massive biomedical datasets, integrating Genomic Data Commons (GDC) and FireBrowse APIs to improve research accessibility. By engineering data aggregation pipelines and implementing Redis-based caching, I successfully reduced API calls from 50K+ daily to just 15K, improving system performance by 70%.

I also optimized PostgreSQL query execution plans, cutting database response times from 2 seconds to under 700ms, and developed interactive D3.js visualizations for genomic data analysis. This experience helped me strengthen my skills in data engineering, API integrations, and scalable data infrastructure, while also deepening my understanding of cloud computing and high-performance computing for research applications. 🚀

Under the mentorship of Professor Haipeng Xing, I worked on developing deep learning models for financial option pricing, aiming to improve accuracy over traditional pricing models like Black-Scholes. Our research involved designing and training neural networks for stochastic modeling, leveraging real-world market data from over 1 million option contracts. By applying adaptive loss functions and hyperparameter tuning, we were able to enhance prediction accuracy by 25% while significantly reducing inference time.

A major focus of this research was on calibrating deep learning models to financial time-series data, ensuring robustness in volatile market conditions. We explored Monte Carlo simulations, reinforcement learning techniques, and volatility estimation models to refine pricing predictions. Additionally, we optimized gradient-based optimization techniques, reducing volatility estimation errors by 20%. This research strengthened my expertise in financial modeling, deep learning architectures, and quantitative analysis, providing valuable insights into AI-driven financial forecasting and market risk assessment.

As a Backend & Data Engineering Intern at Baaamboos For Men, I worked on optimizing database performance, integrating ERP systems, and automating data pipelines to improve business operations. I developed RESTful APIs for seamless data exchange, reducing backend response times by 40%, and implemented secure authentication mechanisms (JWT, OAuth 2.0) to enhance data security.

I also focused on database optimization, improving MySQL query performance by 50% through indexing and query execution plans. Additionally, I developed batch processing scripts in Python to automate transaction handling, reducing data inconsistencies by 30%. By integrating Redis caching, I minimized API load and significantly improved response latency.

This role gave me hands-on experience in backend development, API design, and database engineering, teaching me how to build efficient, scalable data systems. Working in an e-commerce environment, I learned the importance of real-time data processing and automation, skills that have shaped my approach to scalable AI and data-driven applications