Viral Dalal

Around that time, I came across SimPPL through some of my classmates' profiles. Curious, I explored their website and the projects they were undertaking. The kind of research and work they were doing intrigued me immediately. SimPPL's tasks were challenging and very hands on. I decided to take on the task of scraping data from Mastodon and performing data analysis. Much to my delight, Swapneel and Dhara loved my approach, and I got selected.

Soon after joining SimPPL, I was given the opportunity to join the LLMxPsych team. The project's aim was ambitious and fascinating: to see whether large language models could actually mimic the collective judgement of human contributors who write Community Notes on X. What really drew me in was both my keen interest in LLMs and the fact that the project was still in its initial stages, which meant I could have a real impact from the ground up. The team, led by Swapneel and the project lead Aryamonvikram, were still brainstorming research questions and evaluating frameworks for building the multi-agent debate system that would simulate how multiple humans discuss and decide on fact-checking notes.

I quickly took the initiative to scope out existing solutions for multi-agent orchestration. After exploring several frameworks, I realized that none offered the flexibility we needed, so I set out to build the architecture from scratch using LangChain. Leading this technical development was an incredible learning curve—designing and implementing the debate-based system that let multiple LLMs "argue," critique, and come to a consensus just like a real online crowd. This work also pushed me to learn the essentials of scientific research: framing rigorous hypotheses, designing experiments, documenting everything thoroughly, and collaborating in an organized, transparent way. Under Swapneel's mentorship, I learned how to shape a formal research paper and saw first-hand how deep and demanding real research can be compared to my earlier perceptions.

During this project, I learned essential skills that went beyond just coding or analysis. I got introduced to effective project management tools like Notion, the importance of weekly meetings and regular updates, and meticulous documentation of every experiment, tech stack, and framework used. This process not only streamlined my own work but also benefited the entire team. I also learned the formal process of conducting experiments and writing research papers.

My success with the LLMxPsych project culminated in an offer for a research engineer internship, a role that shifted my focus toward more applied engineering challenges. I joined the Newsroom Analytics team to help build 'Axioma', an AI-powered platform designed to give newsrooms the data-driven insights they need to grow.

The platform offered a comprehensive dashboard and provided a deep dive into audience analytics, revealing important insights like not just who was reading, but also where they were coming from (traffic sources like Google and Reddit) and what content kept them most engaged. More powerfully, it used AI to analyze online trends, predict which topics would resonate with specific user segments, and even identify untapped content areas, giving editors a clear compass for their content strategy. The result was one of the most impressive demos I have ever been a part of, a fully functional MVP that brought all these insights to life in a clean, interactive dashboard.

However, this project was also a crucial lesson in market realities. We learned that many newsrooms, our target audience, were facing significant financial pressures that limited their ability to invest in new digital tools, no matter how effective. This real-world business challenge prompted a strategic pivot.

Realizing the difficulties of external markets, a conversation with Dhara shifted my focus inward to a critical challenge facing our own organization. As a nonprofit, SimPPL relies heavily on grant funding to support its work. Swapneel and Dhara were dedicating immense time and effort to this, manually drafting hundreds of applications by drawing from their extensive repository of past proposals.

They had an incredible asset—hundreds of previously successful proposals, but leveraging that knowledge for each new submission was a painstaking manual process. This sparked the idea for a problem solving project: creating an internal AI tool to streamline this effort. I took on the challenge and, in about a month, delivered a fully functional RAG system. The system used SimPPL's own winning proposals as its core knowledge base to draft new, compelling applications, effectively turning their historical data into an active assistant.