Delivering impact with Data & AI
Edinburgh, Scotland
Commercial Director at Blend. Seventeen years helping organisations in the energy sector turn data into real business value - by building the right teams, asking the right questions, and keeping the focus on what actually matters.
About
My career has taken an unusual path. I studied chemistry, started in energy research and analysis at Wood Mackenzie, then spent a decade in progressively senior commercial and strategy roles at TotalEnergies - in Aberdeen and Johannesburg. Along the way I built and led a Digital and Data team of 35 people that delivered over $200M in value in its first three years.
In 2023 I joined Blend as the founding leader of their Energy practice. I now serve as Commercial Director, helping clients across sectors do data and AI right - matching real needs to real solutions, and making sure the value lands.
The thread running through everything I do is a belief that the best outcomes come from honest conversations, well-built teams, and a relentless focus on what works in practice rather than what sounds good in a presentation.
Projects
Independent work outside my day job — building tools and exploring datasets that interest me.
An end-to-end data pipeline and intelligence platform covering the UK Continental Shelf. Pulls monthly field production data from NSTA's public ArcGIS API, scrapes regulatory news and company RNS announcements, tracks Brent crude and share prices daily, then uses GPT-4o to generate structured analyst briefings for the four major independent operators — Harbour Energy, EnQuest, Serica Energy, and Ithaca Energy. Data is stored in Databricks Delta tables; briefings are served through a Next.js web application.
View on GitHub → Live app →A research and experimentation project using Equinor's open Volve dataset — a decommissioned North Sea field — as a realistic testbed for evaluating Databricks (Unity Catalog, Delta Lake, serverless compute) alongside Claude Code with Sonnet 4.6 as an AI development partner. The task itself is an XGBoost classifier that predicts production flow upsets 2–14 hours ahead, using only independent signals — downhole pressure and temperature sensors and topside rotating machinery — rather than the surface flow meters that would make the problem trivial. The investigation surfaced several real data quality and modelling pitfalls: a minimum lookahead gap that had to be enforced to prevent the model from detecting ongoing events rather than predicting future ones; a PI historian sensor failure silently injecting -999 sentinel values into the feature matrix; and a dominant calendar feature (month) that turned out to be a training artefact from the timing of a major shutdown rather than a genuine seasonal signal. The biggest takeaway: you need a production engineer to make a dataset like this genuinely useful. The data captures what happened — but not why a shutdown was delayed, which sensor everyone on the platform knew to ignore, or what was observed before the event showed up in the historian. The tacit knowledge simply does not exist in any dataset.
View on GitHub →An end-to-end drilling analytics pipeline built on 15 million 30-second WITSML sensor records from Equinor's open Volve North Sea dataset. Parsed raw XML logs across 14 wells into a Databricks Delta Lake, then built two complementary analyses: a within-well XGBoost model (80/20 time-based split) that assigns an efficiency score to each drilling hour by comparing actual ROP to what the model expected from the applied parameters; and a formation-level quartile benchmarking study that compares median drilling parameters in the fastest vs slowest 25% of on-bottom rows across 13 geological formations. Key findings: formation type dominates ROP far more than drilling parameters; Skagerrak is WOB-dominant (15 kN, low RPM → 102 m/hr), Ty is RPM-dominant (160 RPM, light WOB → 29 m/hr); and high ROP in deep reservoir formations reflects natural fractures rather than aggressive drilling — the fastest rows in Aasgard and Draupne used less mud circulation than the slowest. The biggest takeaway: you need a drilling engineer to make a dataset like this genuinely useful. The Volve dataset is to an asset what a cave painting is to a hunt — a record of what happened, silent on why.
View on GitHub → View analysis →Contact
Whether you want to talk about data and AI in the energy sector, explore working together, or just have a conversation — feel free to reach out.
niall.rowantree@blend360.com linkedin.com/in/niall-rowantree