George Weale

Software engineer working on agent evaluation and systems

George Weale

Research

I study how to measure and understand complex systems, from AI agents to chemical processes. Here are some of the questions I’ve worked on.

Featured paper Dynamics and prediction

Can four noisy position measurements predict which particles will escape? I tested this in a system with three exits, withholding both the energy and starting region from each predictor.

Short position histories helped at nearby forecast times. Later gains and the benefit of adding an instability measure remained unresolved.

Read PDF 6 pages · 713 KB

Hénon-Heiles Escape Prediction from Noisy Position Histories

Three simulated Hénon–Heiles paths at energy 0.18 wind through the central region before exiting upward in blue, lower-right in gold, and lower-left in coral. Circles mark their starting positions, and faint curves mark potential-energy contours. These are illustrative trajectories from the study, not a chart of prediction accuracy.Three simulated Hénon–Heiles paths at energy 0.18 wind through the central region before exiting upward in blue, lower-right in gold, and lower-left in coral. Circles mark their starting positions, and faint curves mark potential-energy contours. These are illustrative trajectories from the study, not a chart of prediction accuracy.
Illustrative speed
Three recorded simulation paths at the same energy, each reaching a different exit. Faint lines show potential-energy contours. These paths illustrate the system, not prediction accuracy.

Engineering and energy

  1. Which sensors does a distillation column need?

    I studied which flow measurements make a selected set of distillation balance equations solvable. Eight of ten three-sensor combinations identify all five flows in this static model; the study does not establish dynamic plant observability.

    Read PDF 3 pages · 300 KB

  2. Do fault-detector rankings hold up?

    I tested five fault detectors on new runs of a chemical-process simulation. The tested Bayesian alarm rule produced fewer alerts and missed more faults. Its operating rule and the unequal alert burdens limit what that comparison says about the underlying methods.

    Read PDF 6 pages · 349 KB

  3. When prices change, does the best process change?

    I compared four ways to make dimethyl carbonate using a published cost model. Hydrogen and product prices affected profitability, but the direct-urea route kept the lowest break-even selling price throughout the reported range.

    Read PDF 2 pages · 201 KB

  4. Choosing where to reuse heat

    I compared cost and reversible cold-stream heating demand in a heat-exchanger model. Lower demand came with greater utility use and cost among feasible operating conditions, without a supported improvement in capacity success.

    Read PDF 5 pages · 317 KB

AI and software

  1. Predicting whether an AI agent succeeded

    Using 4,448 benchmark runs, I tested whether summaries of an agent’s actions helped predict task success. They added no reliable benefit over ordinary interaction records when both the tasks and agent models were new to the predictor.

    Read PDF 4 pages · 253 KB

  2. Checking whether rewritten code behaves the same

    I checked Python and Java programs against shared test inputs. The checker caught 58 of 60 eligible known-bug cases on QuixBugs; that result applies to the tested inputs, rather than proving the programs equivalent.

    Read PDF 5 pages · 207 KB

  3. Finding what slowed a program down

    I added delays to different parts of a program and tested whether its timing logs could locate them. Comparing each step with its normal behavior identified the delayed step in 119 of 144 tests; missing records and noisy tool processes limited the result.

    Read PDF 5 pages · 216 KB

  4. Strust: comparing COBOL and Java

    The Strust v1 paper describes comparing a COBOL program with its Java replacement under a shared test contract. It includes a ten-case historical ledger and a proposed verification architecture; the preserved records do not establish production accuracy or scale.

    Read PDF 7 pages · 269 KB

Biology and physics

  1. Choosing which lipid nanoparticles to test

    Can published lipid nanoparticle data help choose the next experiment? Across 13,982 formulations, the models did not establish a reliable screening gain over random selection on unseen publications. Comparisons with random splits also changed the candidate pools, budgets, and fitting procedure.

    Read PDF 5 pages · 265 KB
Earlier projects

  • Optimizing MoE Routers

    Compared transformer router designs across latency, expressiveness, and expert utilization.

    Read PDF

  • Time Series Prediction

    Modified the LSTM-FCN architecture with squeeze-and-excitation blocks and compared it with replication baselines across time-series datasets.

    Read PDF

  • Pico Size Chatbot

    Built and trained a small GPT-2-style chatbot on AWS and a university compute cluster.

  • MealMapper - Columbia Startup Lab

    Built a meal-tracking system with PyTorch and YOLOv8 for food recognition and nutrition estimates, then deployed it on Google Cloud for sports nutritionists.

  • Senior Design 2 - UCSB

    Designed and evaluated a carbon-negative dimethyl carbonate plant, including equipment, economics, and the effect of European carbon credits.

    Read PDF

  • Senior Design 1 - UCSB

    Ran a techno-economic analysis of a carbon-neutral steam ethane cracker, covering the reactor, distillation train, and heat-exchanger network.

    Read PDF