Data, code, and public goods
I build tools and datasets for answering questions that conventional, off-the-shelf data can’t. My work spans two kinds of alternative data—text and geospatial data—and brings together machine learning, natural language processing, and geospatial analysis and remote sensing.
Mapping India's solar expansion
Trace the expansion of solar farms in nine Indian states over the last decade, identified using satellite imagery and deep learning.
The Indian energy transition is critical for both conservation and economic independence. Several Indian states rapidly expanded solar farms in the early 2020s, a transition that traditional administrative data cannot capture. This dashboard maps solar farms across India and tracks when they were built.
Statistical packages
Open-source software contributions for people who use econometric methods.
Python + Rust
fastLP
Fast panel local projections for applied macroeconomists, with a beginner-friendly Python API and a high-performance Rust core.
Caches and reuses the same calculations across horizons whenever possible instead of rebuilding the same matrix for each horizon.
pip install fastlp-py
Stata
longview
A Stata command for plotting cumulative treatment effects from difference-in-difference event-study designs.
Aggregate multiple estimates across time; useful when the research question focuses on long-run consequences of a policy or intervention.
net install longview, from("https://raw.githubusercontent.com/ritwizsarma/longview-stata/main/")
Public goods & resources
Python: requests
CAG audit report scraper
A scraper that makes the Comptroller and Auditor General of India’s extensive archive of audit reports easier to collect and study as text data.
Python: Playwright
Plot-level land values
Repository of code for scraping very high resolution datasets of land values from public registration systems.
Data science projects
Opportunities where I built my technical skills.
Text data
In progressMonetary policy sentiment from text
I fine-tuned a BERT-based model to classify hawkish/dovish sentiment at the sentence level, then apply it to the full archive of Federal Reserve transcripts dating back to 1990.
Text data
ConcludedJudicial text classification
Deep learning and preprocessing pipelines for a unique judicial text-classification problem at XKDR Forum.
My work included training and fine-tuning on several architectures, including LSTMs and CNNs, and comparing the efficiency of RNN-based pipeline to an LLM-based approach.
Intern forAjay Shah, Siddarth Raman.
Text data
In progressGazette of India corpus
A massive corpus assembled from multiple Gazette of India archives, with ongoing NLP work to turn these public notices into structured, usable data.
Please reach out to me if you'd be interested in contributing!
WithAnkit Bhatia.
Geospatial data
In progressGeospatial dataset of bridges
A novel dataset of Indian bridges with geospatial and temporal markers, built for a difference-in-differences study of infrastructure and economic activity.
WithAnkit Bhatia.