The data
was already
there.

A routine startup.
Nothing unusual.

The raffinate splitter tower was being restarted after maintenance. It was a procedure the operators had run before. The control room was staffed. Instruments were live. The safety systems were in place.

What nobody saw, across the three separate control rooms monitoring different sections of the unit, was what the data looked like when you read it all together.

The level in the tower was rising. It had been rising for hours. The level indicator — a float-based gauge — was reading incorrectly, showing the vessel as partially full when it was dangerously overfilled. A second high-level alarm had been out of service for years. The operators were working from the instrument they had, trusting it the way you trust an instrument that has always been there.

The tower overflowed.
The drum filled.
The vapour cloud ignited.

At 1:20pm, the raffinate splitter overflowed into the blowdown drum. The drum — designed for liquid, not the volume it now received — also overflowed. Hydrocarbon liquid and vapour discharged from the stack into the atmosphere around the unit.

A contractor's pickup truck, engine idling in a prohibited zone nearby, provided the ignition source.

All fifteen fatalities were contractors working in or near temporary trailers positioned too close to the unit — a placement that violated BP's own siting guidelines, documented in their own safety management system.

The investigation asked
the harder question.
Was this foreseeable?

The US Chemical Safety Board spent two years reconstructing what happened. Their conclusion was not that the technology failed, or that the operators were negligent, or that the safety system was absent.

The instruments were live. The maintenance records showed the level gauge history. The operating procedures documented the startup sequence. The incident record — from previous upsets at the same unit — contained clear warning signs that this configuration was vulnerable.

Every piece of information needed to prevent the explosion existed somewhere in the plant's data. It existed in three separate control rooms. It existed in maintenance logs. It existed in the safety management system. It existed in previous near-miss records.

What did not exist was a single view of what all of it meant together, in real time, interpreted by someone who understood the combination.

One question changes
everything.

After Texas City, after Buncefield that same year, after Piper Alpha seventeen years before it, after Bhopal, after every major process industry incident that has been investigated thoroughly enough — the finding is structurally the same.

The data was there. The connections between it were not.

The question is not: how do we prevent explosions? The question is narrower and more honest:

In Texas City: almost certainly yes. The level was rising for four hours. A competent engineer with full visibility of what was happening across all three control rooms, cross-referenced against the maintenance record on that level gauge and the known vulnerability of that blowdown configuration, would have stopped the startup.

The problem was never knowledge. The problem was that no single person had all of it, in one place, at the moment it mattered.

That question
is what
StarFarer answers.

StarFarer is not a chatbot. She is not a search engine. She is not a document management system.

Aby — StarFarer's reasoning intelligence — reads your P&IDs, your incident records, your maintenance history, your operating procedures, and your regulatory obligations simultaneously. She holds your plant's entire knowledge state at once, the way a senior engineer would if they had unlimited time and perfect recall. She tells you in plain English what the combination of this equipment state, this operating mode, and this maintenance backlog means for your risk profile — before something happens.

Every finding is cited to its source. Every answer carries a confidence score. When the evidence is thin, she says so. A confident wrong answer in process safety is the most dangerous possible output. The whole platform is engineered against it.

She is on duty at 3am. She is on duty over Easter. She does not miss a shift, does not have a bad day, and does not leave for a better opportunity in another city.

Built inside an
engineering practice.
Not beside one.

StarFarer was not built by AI researchers who then learned process safety. VyroNyx is a process safety and energy engineering firm. We deliver HAZOP studies, LOPA analyses, safety cases, and regulatory submissions. StarFarer was built inside that practice — on real studies, real P&IDs, real incidents, real regulatory pressure — before she was offered to anyone else.

The knowledge base is hand-encoded by senior engineers. It is not generated by a model trained on the internet. It is reasoning chains — the kind your best process safety engineer carries in their head after twenty years. That is slow to build and impossible to shortcut. It is also exactly why it works.

Our paper on propagation containment in integrated solar-BESS-electrolyser architectures was accepted for oral presentation at IChemE Hazards 36. That is the peer-reviewed frontier of process safety for the energy transition. StarFarer reasons from it.