Sai Krishna Rallabandi

Sai Krishna Rallabandi works on AI systems for financial question answering and investment-report generation. He describes himself on his personal website as Director, Data Science at Fidelity Investments. His collaborative research examines how models carry out numerical calculations and how writing agents can use feedback to improve reports.[2][3][4]

Rallabandi also develops independent personal projects exploring how AI systems gather context, act, and remain accountable. In July 2026, Latent Space’s finance-track coverage associated his participation with group-chat and wearable agents, focusing on memory, permissions, and prompt-injection defense.[2][1]

Earlier, his 2025 research with Fidelity colleagues addressed two concrete problems in financial AI: generating accurate calculations and refining written investment reports.[3][4]

With Anushka Yadav, Parag Pravin Dakle, and Preethi Raghavan, Rallabandi coauthored a January 2025 study that separated financial question answering into intermediate calculations and a final answer. The team reported that including those calculations improved numerical accuracy, then compared traditional direct preference optimization (DPO), iterative DPO, and length-regularized DPO.[3]

The experiments exposed a tradeoff: traditional DPO eliminated many parsing errors but sometimes made simple calculations unnecessarily complex. One model generated [divide(1,B), multiply(A,#0)] where divide(A,B) would suffice. The authors linked this verbosity to reduced parsing accuracy and reported that length-regularized DPO reduced verbose intermediate calculations while improving numerical accuracy across the tested models.[3]

For the Earnings2Insights shared task, Rallabandi coauthored research with Nikhil Kohli, Parag Pravin Dakle, Khyati Morparia, Ojas Raundale, and Preethi Raghavan. The team compared four ReAct-based report-generation frameworks: two single-agent systems powered by a Writer agent and two modular multi-agent systems incorporating Feedback agents for iterative refinement. A six-metric LLM-as-a-Judge framework ranked the generated reports.[4]

The team reported that its ensemble achieved average financial accuracy of 0.571, placing third, and an average Likert score of 5.90, placing second, in the shared task’s human evaluations. It also placed first in readability and second in next-day prediction accuracy. These results measured the report-generation ensemble in the shared task, rather than investment returns.[4]