George Weale

Work

Now

Google Cloud AI Software engineer

I work on evaluating AI agents and improving their reliability, speed, and running costs.

Agent Development Kit

Strust

Strust

I built Strust to run COBOL programs and their Java replacements on the same tests and compare the results.

Project website

Research

2026

• Predicting whether an AI agent succeededPDF

  State-change features helped one fitted procedure within the original agent cohort, but that gain did not transfer convincingly to the new held-out agent. The study separates learner interaction from reusable predictive value.

• Checking whether rewritten code behaves the samePDF

  External repair candidates show why a disagreement is not automatically a detected defect. Reviewed contracts and concrete witnesses separate real failures from tolerance and mutation differences.

2025

• Finding what slowed a program downPDF

  Timing-based fault rankings can collapse when a clean span shifts. New controlled runs expose the timing boundary, partial repairs and the limits of an alert threshold learned from only two controls.

• Strust: comparing COBOL and JavaPDF

  Two correct implementations can agree in the wrong replay state. Executed transaction histories expose four such cases, while a cached-output baseline shows why slicing adds work for this unchanged-pair validation task.

• Choosing which lipid nanoparticles to testPDF

  I compared models on matched candidate pools and testing budgets across unseen publications. Tie handling and uncertain readout directions limit the screening claim; source spreadsheets reconstruct 60 retained candidates.

2024

• How far ahead can we predict a chaotic system?PDF

  With the same four noisy positions and known energy, physical state inference predicted near-term escape much better than the learned models. Independent integration also exposed errors in the later numerical event labels.

• Which sensors does a distillation column need?PDF

  Two measurements can distinguish reflux from feed-composition changes yet estimate them very poorly. A component model, matched-noise fits and sensitivity directions explain why instrument placement matters.

2023

• Do fault-detector rankings hold up?PDF

  I repaired a Bayesian detector’s recurrence and checked it against exact small-case calculations. Replaying the coherent detector on matched test runs retained its large recall deficit, with unequal realized alert burdens.

• When prices change, does the best process change?PDF

  Published break-even prices and investment values imply incompatible production factors under the stated accounting boundary. The audit identifies that inconsistency without guessing a corrected process result.

2022

• Choosing where to reuse heatPDF

  A lower heat-demand score can still require more supplied heat and cost. I checked that reversal across alternative dispatches and independently reconstructed the retained designs’ training feasibility.

Research details and figures

Writing

• Oct 2026The Mistakes an Agent Doesn't Ask About
• Jul 2026Why Requests Can Slow Down Without Higher CPU Usage
• May 2026How Work Changes Around a Tool

All writing

Education