WellPilot

WellPilot answers an engineer's questions about oil well performance from the field's own data and reports, and has to show its work. Engineers who look after oil wells spend much of their time pulling data out of production databases, PDF reports and spreadsheets before they can say why a well changed. In WellPilot, an agent on Muse Spark Open plans its own steps and calls real analysis tools over the field's data: decline curve fitting, forecasts with ranges, event detection, well work from daily reports, production logs and the 2005 development plan. A separate checker, plain code with no AI, reads every number, date and quote in the answer against the tool result it cites and sends the answer back until all of them are found. Every number in the final answer links to its source and to the code that produced it. The demo hides all data after 2010. The agent finds that well F-12's pressure reading went to zero because the gauge failed, not the reservoir, and its forecast for 2011 came within 6% of what the well produced (359,000 against 340,000 m3; the textbook whole-history fit was 102% too high). On 25 past forecasts with the future hidden, its median error was 14%, about even with the best fixed fitting window, which is known only in hindsight, and half the textbook fit's 28%. Asked what to do with the same well in 2014, it names a 14-metre stretch where water could be shut off, warns that a seal there would also cut most of the oil, and asks for a fresh survey first. A question takes about a minute and a few tens of cents; by our estimate the same answer takes an engineer a day or two. Live demo: https://wellpilot.fly.dev (recorded answers open instantly, no sign-up; live questions are rate-limited) Built solo during the hackathon, from the first commit on Oct 26. Data: Equinor's open Volve field dataset, under the Equinor Open Data Licence; not affiliated with or endorsed by Equinor.