Research

Projects at Stanford, IIT Bombay, ISTA and Stanford GSB, newest first.

Foundation models for relational data

Architecture and pretraining improvements to the Relational Transformer, so that few-shot in-context prediction on relational databases is competitive with task-specific fine-tuned models on RelBench forecasting and autocomplete tasks.

LLM tool-use agent for DevRev

Fine-tuned agent that turns conversational queries into a graph of tool calls, adapting to the tools available. Built a synthetic dataset with ReAct prompting over ~100 strongly typed domain tools, and a graph-based correctness metric that ignores the order of independent calls. Fine-tuned 7B–13B open-source models with QLoRA on masked tool and argument prediction, benchmarked against GPT and Claude, and added a lightweight high-recall embedding retriever to narrow the tool search space.

Causal time-series retrieval

Retrieval module for the TextSpaceTime time-series foundation model. It greedily adds the database series with the highest transfer entropy against the residual of the target given the series already retrieved, discounting autoregressive information in the target, after Granger causality.

Generator-assisted feature acquisition in batch

One-shot acquisition of a feature subset under a cost budget, as in diagnosis with selective tests. Routes each instance to a data cohort with locality-sensitive hashing, acquires that cohort's chosen features and imputes the rest with a VAE. Beat the state of the art on 4 benchmarks.

blog | AAAI'24 | arXiv

Multi-strategy Moran processes

Fixation probabilities for Moran processes on graphs with three competing strategies. Found a martingale construction that generalizes beyond two strategies and gave new bounds for very low and very high mutation probabilities, presented as an internship thesis.

Sentiment congruence in advertising

Field experiment with the content platform VDO.AI on how emotional congruence between a video and the ad served with it affects engagement, controlling for day, time, region and the video's base sentiment. Incongruent ads converted better.

SSRN