Projects

Automation runs safety-critical processes well inside their design envelope. Abnormal situations still fall to human operators, and their mismanagement is a leading contributor to process-safety incidents. Full autonomy in other domains usually leans on a safe stop. A chemical plant does not have that recourse because stopping is itself a hazardous and expensive operation. This project asks whether a general-purpose large reasoning model (LRM), with no task-specific training and only the information available to an operator, can manage abnormal situations at run time, provided it acts through an interface that cannot express an unsafe command. ...

Dynamic process models are increasingly consumed by automated agents rather than by control engineers reading source code. These agents include data pipelines that assemble training sets, reinforcement-learning algorithms, and language models. Each depends on process meaning that a numerical interface does not expose, such as what each variable represents, which signals are available online, and which limits trigger a shutdown. Today that meaning lives in source code and informal convention, so recovering it is error-prone and does not transfer across a model’s many independent implementations. ...

Book Condenser produces a shortened reading edition of a nonfiction book. Its premise is narrow. The goal is to read less of the book without replacing the book with a summary. Rather than asking a model to rewrite the content, it identifies and keeps, word for word, the passages needed to preserve the central question, thesis, key terms, argument, evidence, and conclusion. What it does Extractive, not generative. Retained text stays word for word from the source. The model decides what to keep, not how to phrase it. Argument-driven selection. Passages are chosen against an analytical map of the book, and essential propositions are kept together with their supporting evidence. Disclosed transitions. When a bridge is needed for continuity, it is italicized and labelled, and it is never presented as the author’s prose. Traceable output. It renders a tablet-optimized PDF (plus Markdown and optional DOCX) with omission markers, and produces audit artifacts documenting the selection decisions. It accepts EPUB, PDF, DOCX, TXT, and Markdown, with configurable target length ratios reported against retained source words only. ...

Selective hydrogenation units remove trace methylacetylene and propadiene (MAPD) from a propylene stream. The more selective way to dose hydrogen is to ratio it against the measured inlet MAPD concentration. That strategy requires a continuous and trustworthy MAPD signal at every control interval, and the online gas chromatograph (GC) that provides it is intermittent and occasionally unreliable. A soft sensor intended for closed-loop use cannot be judged as a standalone predictor. Low aggregate error and near-nominal interval coverage can still hide local reliability loss during operating-state transitions, which is exactly when a controller can least afford a bad measurement. The soft sensor is therefore treated as part of a reliability-aware measurement layer that decides, at each step, whether to trust the analyzer, rely on the model, or route the controller to a conservative fallback. ...