Agentic AI for Household Finance Research
Demo recordings from the NBER SI 2026 talk — real sessions, recorded ahead of time, unedited. Slides and materials: github.com/paulgp/nber_hf_2026.
One prompt → the live poll you just used
A fresh Claude Code session builds the audience poll (Cloudflare Worker + Durable Object), tests it end-to-end — including a 40-vote concurrent burst — and writes the deploy steps. Recorded 2026-07-22, one take. ~7 min of work, playing at 2.5x speed.
Space = pause · ←/→ = seek · press ? in the player for all keys.
The 2018-2024 HMDA build
A three-sentence prompt plus a staged context/ folder of API field notes. The session pulls seven years of Ohio HMDA data in parallel, centers each year's loan bins on that year's conforming limit, and produces the spike-follows-the-limit figure used in the slides. Recorded 2026-07-23, one take, verbose (the SQL and Python are on screen). ~13 min of work, playing at 2.5x speed.
Space = pause · ←/→ = seek · press ? in the player for all keys.
New bunching estimator, cross-checked
A clean directory containing one committed data table. The session implements the Kleven-Waseem/DeFusco-Paciorek notch estimator from a spec, bootstraps the standard errors, then reconciles itself against the CRAN bunching package — discovering on its own that the package's kink-mode statistic nets the missing mass against the bunching mass, and driving the two implementations to agreement in every year. Recorded 2026-07-23, one take, verbose. ~20 min of work, playing at 2.5x speed
Space = pause · ←/→ = seek · press ? in the player for all keys.
Writing a model on top of the empirics
Prompt with the modeling choices (log utility, seven-year duration, partial equilibrium) are stated as fixed; the agent executes the model inversion, re-derives one knob (D = 5) without editing the committed code, runs the 20%-limit-cut counterfactual, and closes by saying what was whose call. Recorded 2026-07-23, one take, verbose. ~5 min of work, playing at 2x speed.
Space = pause · ←/→ = seek · press ? in the player for all keys.
Two GitHub issues → two reviewed PRs, no human in the middle
The coauthor loop as infrastructure: two real issues filed on the repo (add an inflation-adjusted column to an estimates table; fix a y-axis). The session reads them with gh, fixes each on its own branch — including discovering that one figure must be rebuilt from a committed table because the raw data isn't in the clone — opens two PRs with verification described, and comments on the issues. The human appears only at the endpoints: filing and reviewing. Recorded 2026-07-23, one take, verbose. ~12 min of work, playing at 2.5x speed.
Space = pause · ←/→ = seek · press ? in the player for all keys.