Twice in Three Years: Two AI Phase Shifts
ChatGPT in 2022, Claude Code in 2025: first the machine learned to understand us, then it learned to act for us. On what those two shifts changed, and why I expect a third.
Fabio Concina
Head of AI & Analytics at kwantis
About Me
Engineering leader with 10+ years in software and data, focused on taking AI from prototype to production in demanding industrial settings. Currently leading an 8-person AI & Analytics team at kwantis, a ~40-person SaaS company whose product, ID3, runs in production for major energy operators including Eni, Saipem, TotalEnergies, and Weatherford.
Experience
kwantis
Milan, Italy ยท 2015 - PresentHead of AI & Analytics
2025 - Present- Lead the 8-person AI & Analytics team (a focused group within the wider engineering org), owning ID3's AI capabilities end to end, from research to production.
- Shipped LLM and agentic features into ID3, including generative AI for automated drilling-report (DDR) data capture that saves crews ~40 minutes per shift, deployed in production across 4 rigs and 150+ rig-days at 100% IADC compliance.
Manager, AI & Analytics
2020 - 2025- Grew the engineering organization from 5 to 20 and owned the design of ID3's analytics layer.
- Designed an AI anomaly-detection pipeline for data-quality assurance in surface logging.
Software Engineer
2015 - 2020- Designed and built the first version of ID3, shaping its architecture and early direction.
- Deployed on-site at clients including Eni, Repsol, and Enel Green Power to integrate ID3 into their operations.
Education
MSc in Statistics and Actuarial Science
Contributed to the R ChainLadder package as part of thesis research on stochastic reserving methods (thesis, PDF).
BSc in Mathematical Engineering
Final presentation on biological waves applied to epidemic spread and control (PDF).
Publications
Talks
Projects
Forecasting rate limits in claumon
claumon is a dashboard for Claude Code that predicts when you'll hit your usage limits before you actually do. Behind the gauges is a small Bayesian model that simulates how usage can evolve until the window resets, then reads off where it lands and when it would cross a threshold, each with an 80% credible interval.