LSMS Agent

Ask in plain English. Get answers from LSMS-ISA household survey data across 8 African countries — variable discovery, panel merges, summary stats, regressions, plots.

Open the app → View the code on GitHub

What it covers

Burkina Faso Ethiopia Malawi Mali Niger Nigeria Tanzania Uganda

30 survey rounds. 76,775 variables across 2,712 data modules. All harmonized into a single canonical layout and queryable from natural language.

Ask it things like

"Which education-related variables exist in Tanzania round 2?"

"List all rounds available for Nigeria."

"Load the Ethiopia 2013 household section A and show me the first 10 rows."

"What's the distribution of household size in Uganda 2019?"

"Merge Tanzania 2008 and 2010 household modules on hhid and show the row counts before and after."

How it works under the hood

Variable catalog built from Stata metadata and CSV headers — searched by keyword, with optional vector search for synonym recall.
Python sandbox per session — pandas, numpy, matplotlib, statsmodels pre-loaded; state persists across turns.
Claude Sonnet 4.6 with tool use plans and writes the analysis code; the user only sees results.
Chainlit chat UI renders tables and plots inline, can export a session as a runnable notebook, and can auto-draft a reusable crosswalk YAML from a successful multi-round analysis (researcher names the concept, reviews, and commits).

Access

The chat is private. Depending on the deployment mode, access uses either a shared research-group password or Hugging Face OAuth for per-user identity. Reach out if you need access.