Pablo Zavala (www.zavalab.com)
Hi! I am passionate about the uses of technology for human capital development. My areas of professional and academic interests are agentic technologies, quantum computing, public policy, and AI governance. I have been working on agentic calibration for long-horizon investments. These are some of my public repositories:
hierarchical-agent-governance-evals
Agents now hold operational authority: they merge code, allocate budgets, direct other agents. This program tests whether an agent overreaches that authority or declines authority it legitimately holds. Six action-level probe families put agents in those situations; blind AI audit panels grade the outcomes.
The methodological finding mattered as much as the behavioral one. Automated graders false-flagged compliant behavior — I caught and corrected about a dozen instances by hand before reporting any result. The effect that survived scrutiny: Fisher one-sided p = 0.0098 — a behavioral divergence rather than a constitutional failure.
The same standard applied to AI in economic decisions: pre-register, report nulls as nulls, state the tradeoffs.
- safe-market-universes — a pre-registered null result: the model's own confidence rationed scarce human review no better than chance, regret 0.176 vs 0.191 for random allocation. A hand-coded evidence rule reached 0.091.
- Macroeconomic-Simulation-Netlogo — an agent-based model of labor-market adjustment to AI automation. Changing the policy regime alone moves peak unemployment from 14.3% to 3.6%, with workers and random seed held fixed.
- RAG_Eval — an evaluation harness where reranking lifts context precision from 69.0% to 86.6%, while exact match drops. A system can look worse on string overlap and be better grounded. You have to measure both.
- Strike-Team — Pittsburgh's AI investment ecosystem mapped at the Block Center, CMU: 133 active AI firms, USD 6.3B cumulative funding (July 2025 snapshot), with the full report committed.
MS, Public Policy & Management (Data Analytics), Carnegie Mellon University, May 2026 — Highest Distinction. Accelerated MA with Honors, University of Chicago. BA, summa cum laude, Soka University — 4.0 GPA in every economics course.
Founder, NUDG — an AI Resource OS venture in formation, built for proof receipts and human-gated execution: nudgai.com
Open to full-time roles in AI evaluation and research, data science, and AI policy.
Portfolio: zavalab.com · linkedin.com/in/pablo-zavala