Pablo Zavala · AI Safety Evaluation · Research Engineering

Evidence-backed AI Evaluation Projects · Pablo Zavala

Public evaluation repositories, restricted-data research summaries, and product prototypes organized by inspectable artifacts and measured limits.

Project Index

Authority Calibration

Proof
Public pilot for agent authority use: disclosure observed in 14/14 and 12/12 suppression probes; self-demotion in 9/9 and 8/8 supplied-rule trials; 0 observed matched firebreak inversions; separate Codex probe found 6/19 assurance weakening
Evidence
Public reproducible repo
Capability and evidence frontier
Pilot-scale evidence; rare failures remain unmeasured.

AI agents increasingly run pipelines, allocate resources, and coordinate other agents. This project tests whether they use delegated authority correctly: neither exceeding their mandate nor refusing authority they actually hold.

Safe MarketUniverses

Proof
120 episodes; confidence regret 0.176 vs random 0.191
Evidence
Public reproducible benchmark
Capability and evidence frontier
The benchmark tests allocation in a compact finance-style environment; broader oversight settings need separate validation.

In a 120-episode oversight-allocation benchmark, confidence-based routing reached regret 0.176 versus 0.191 for random allocation, too small to serve as an oversight-triage signal.

RAG Evaluation Lab

Proof
86.6 percent context precision after rewriting and reranking; exact match fell on the full 918-query split
Evidence
Public evaluation harness
Capability and evidence frontier
Grounding improved while deterministic exact match declined, so the result is a tradeoff.

An evaluation harness comparing baseline and reranked retrieval-augmented generation pipelines with RAGAS and SQuAD metrics on the Mini Wikipedia corpus. The reranked pipeline reaches 86.6 percent context precision but loses exact match.

DonorsChoose Funding Risk

Proof
ROC AUC 0.757 on 185,000+ held-out classroom projects
Evidence
Public analysis repo
Capability and evidence frontier
The model is a policy triage aid; deployment as a funding decision system would require additional validation.

A model that flags DonorsChoose classroom requests most at risk of going unfunded, so limited reviewer attention can reach under-resourced schools first. The fairness audit reports unequal error rates across school poverty levels as part of the deployment analysis.

SQLite Privacy Boundary

Proof
Six boundary self-checks gate CI: determinism, a k=5 distinct-text floor, a read-only SQL authorizer, redaction at rest, a hash-chained audit, and a private-table-free extract
Evidence
Public reproducible repo
Capability and evidence frontier
The controls stay structural rather than semantic; pattern redaction misses novel self-identifying text, and the floor counts distinct text rather than people.

A dependency-free prototype releases structure from a free-text SQLite dataset without exposing raw rows or private base tables. Pattern redaction cleans text before storage, a distinct-text floor of k=5 gates every public topic and aggregate cell, and SQLite's authorizer denies writes, private tables, and sensitive columns on each query.

NUDG

Proof
Authorization, constraints, verification, and receipts for agent-run work
Evidence
Founder system, bounded public claims
Capability and evidence frontier
Public visuals explain the system model; live product claims require separate proof packets.

NUDG is a CMU AI Venture Studio project for controlling how agents use real resources. It replaces broad agent access with stepwise proposal, authorization, execution, verification, and receipt layers.

AI Investment Mapping

Proof
$10B+ mapped across 11 metros; Pittsburgh: $6.3B across 133 firms
Evidence
Restricted data, public aggregate summary
Capability and evidence frontier
Company-level records and maps stay private; the public page shows aggregates and methods evidence.

A Block Center project mapping more than ten billion dollars in public and private AI investment across eleven metropolitan economies for regional AI-readiness research. The Pittsburgh slice covers 6.3 billion dollars across 133 firms.

AI Workforce Simulation

Proof
14.3 percent vs 3.6 percent peak unemployment under paired policy regimes
Evidence
Public simulation repo
Capability and evidence frontier
Mechanism demonstration in a small simulated labor market, scoped away from macro forecasting.

An agent-based NetLogo model of a small labor market adjusting to AI automation. With identical workers, geography, and random seed, peak unemployment reaches 14.3 percent under a tech-driven policy regime versus 3.6 percent under a human-centric one.

Heard.now

Proof
Synthetic public sample with 7/7 privacy and integrity checks
Evidence
Private pilot, synthetic public artifact
Capability and evidence frontier
Public visuals use synthetic text so community messages stay private.

A civic-listening pilot with Professor Jordan Usdan of Heinz College that stores raw input privately and publishes privacy-checked extracts. The public synthetic sample demonstrates a 7/7 verification path while keeping community messages private.

CMU Event Compass

Proof
Claude vision extracts flyer cards; backend-free static export runs with a sample board
Evidence
Live static demo
Capability and evidence frontier
The demo verifies the sample-board extraction workflow; live campus coverage remains outside scope.

A prototype that turns a photo of a campus poster wall into structured, personalized listings. Claude vision extracts one listing per flyer, a deterministic in-browser ranker orders the results by chosen interests, and the app ships as a backend-free static export.

DemFlex

Proof
Hour-by-hour cashflow model with thermostat, solar, and battery portfolio search
Evidence
Request-only capstone artifact
Capability and evidence frontier
The public artifact shows the function flow; full capstone materials are private.

A Streamlit planning tool for ERCOT demand response. It compares thermostat, solar, and battery portfolios hour by hour through benefit-cost cash-flow analysis.

Cybersecurity Anomaly Detection

Proof
Frozen-threshold holdout: recall 0.9000 (162 of 180 attacks; 95% CI 0.85 to 0.94) at FPR 0.0194, ROC AUC 0.9714, from an Isolation Forest fit on benign-only traffic
Evidence
Public reproducible repo + executed notebook
Capability and evidence frontier
Precision at the chosen threshold is 0.3189 at simulated 1% prevalence, the malicious class holds a single botnet-staging campaign on one honeypot host, and three dominant features leave the detector spoofable.

A Carnegie Mellon security-analytics project that fits an Isolation Forest on benign-only Linux kernel traffic from the BETH honeypot dataset, then grades it under a frozen threshold. On an 18,000-event holdout at simulated 1% attack prevalence, the detector catches 90% of intrusions (162 of 180) at a 1.94% false positive rate, with ROC AUC 0.9714. The entry audits its own headline: precision sits at 0.3189, so roughly one alert in three flags a real attack, and three dominant features hand an attacker a spoofing playbook, which keeps the detector one layer in defense in depth.

LA Housing Hedonics

Proof
Across roughly ten tuned learners, random-CV R-squared 0.621 falls to 0.431 under spatial-block validation, and a leave-one-region-out spectrum sends tree-ensemble skill toward zero; Double ML estimates a +29.8% direct coastal premium (95% CI [0.16, 0.33] log points) and shrinks the naive private-school gap to +0.7%.
Evidence
Public reproducible repo + working paper
Capability and evidence frontier
Causal readings hold only under ignorability, the coastal estimate rests on 100 treated homes, and third-party terms keep home-level rows out of the repository, so a full rerun starts from the cleaned table the committed ingestion code rebuilds.

A University of Chicago Booth big-data study, co-authored with Will Sigal, that prices 3,804 Los Angeles County single-family homes from structural and neighborhood features, then audits its own headline. A tuned stacked ensemble reaches out-of-sample R-squared 0.621 under random cross-validation and falls to 0.431 when scoring moves to held-out geographic blocks; on the separate leave-one-region-out spectrum, tree-ensemble skill decays toward zero while the humble regularized linear model holds up best across space. Cross-fitted Double/Debiased ML puts the direct coastal premium near +29.8% per square foot, while the naive private-school premium collapses to a statistical zero.

Last-Mile Delivery

Proof
In simulation, risk-aware MILP assignment with 240 agents (base risk weight alpha = 1) reached 94.8 percent on-time performance at 89.4 percent coverage, against a 48.8 percent observed baseline on the same synthetic delivery data.
Evidence
Public repository (committed notebooks and report)
Capability and evidence frontier
Every headline rate comes from simulation on a synthetic Kaggle delivery dataset; the report flags capacity calibration as the fragile assumption, and the dataset itself stays outside the public repository pending license review, with a README pointer to the download.

A predict-then-optimize framework for last-mile delivery, built with Santiago Enríquez as a balanced two-person team for Carnegie Mellon's 94-867 Data to Action course. Gradient-boosted models forecast a central estimate and an 80th-percentile bound for delivery time on a synthetic Kaggle dataset of Amazon-style delivery records; the triage MILP screens orders on the tail forecast, while the simulation MILP prices each order by a classifier's risk of missing its conformally set service promise. In day-by-day simulation, the risk-aware assignment reached 94.8 percent on-time performance at 89.4 percent coverage against a 48.8 percent observed baseline.

Dow Reddit Volatility

Proof
Across 3,183 headline words at 10 percent FDR, returns keep one survivor (damn, p = 1.0262e-05) versus 12 words for volatility; adding previous-day volatility lifts the lasso from zero selected words to in-sample R-squared 0.262.
Evidence
Public notebook repository
Capability and evidence frontier
In-sample evidence from one executed notebook on a dataset ending July 1, 2016; the source CSVs stay outside the public repo for license reasons, so a rerun starts from the Kaggle download plus the derived word files the README documents.

A Carnegie Mellon big-data coursework notebook that asks whether words in daily Reddit news headlines predict the Dow. Across 3,183 filtered words, marginal regressions with Benjamini-Hochberg FDR control leave exactly one return-predictive word, damn, a survivor the notebook reads as minimal predictive ability rather than discovery; volatility keeps 12 words and carries genuine persistence. Adding previous-day volatility lifts the lasso from zero selected words to an in-sample R-squared of 0.262, a compact lesson in multiple testing and confounding.

Poker Opponent Modeling

Proof
Adaptive poker bot with a documented study workflow: posterior opponent routing, packaged bucket-EV scoring, uncertainty-gated turn search, and manifest-driven duplicate-deck backtests with frozen release snapshots.
Evidence
Public submission repository
Capability and evidence frontier
Evidence rests on the committed bot and the documented method; the duplicate-deck study scripts that regenerate the comparisons stay outside the public tree, a public headline benchmark against the live tournament field remains pending, and the opponent models stay heuristic and statistical.

A tournament poker bot built for the CMU Data Science Club competition, aimed at decision-making under uncertainty against varied and adversarial opponents. The bot reads opponent tendencies from live betting signals, routes among specialist strategies through a bandit-style selector, and spends heavier turn-street search only on high-uncertainty spots. A manifest-driven backtest harness with duplicate-deck reconciliation and frozen release snapshots records every promotion decision against fixed opponents and score weights.