Brightcon 2026, hackathon and courses

Highlights / Discussion Points

• A global physical uncertainty-aware oil Supply-Use Table (SUT) compiled through an AI-assisted scientific workflow.
• Human input was limited to prompting, review, and modelling judgement; AI generated the code.
• Demonstrates R/targets as a reproducible backbone for agentic data engineering.
• Discusses strengths, failure modes, memory management, and assumption tracking in AI-assisted database construction.
• Invites discussion on transparent, auditable AI-assisted sustainability data infrastructure.

Concise Description

LCA and sustainability modelling increasingly depend on complex, multi-source data pipelines. These pipelines require not only domain expertise, but also extensive software engineering: source adapters, mapping between classifications, validation routines, uncertainty handling, documentation, and reproducible execution. This presentation uses the compilation of a global physical oil supply-use table as a case study in agentic AI-assisted data engineering.

The underlying data product represents the 2022 global oil supply chain at country level, covering extraction, refining, bilateral trade, and sectoral end use in physical units. It includes detailed crude and refined-product categories, uncertainty metadata, validation checks, and a Bayesian balancing step. The product is relevant to LCA because petroleum flows are often embedded in background databases, transport systems, petrochemical supply chains, energy scenarios, and product carbon footprints, yet the underlying physical data and reconciliation choices are rarely transparent.

The workflow was implemented in R using targets and reproducible environment management. However, the distinctive feature is the process: code was produced through agentic AI, while the researcher acted as modeller, reviewer, and prompt designer. The talk will examine where AI was highly effective, including rapid scaffolding, repetitive data-wrangling code, schema consistency, automated reporting, and debugging support. It will also discuss limitations, including source misinterpretation, overconfident assumptions, increasing forgetfulness and laziness, fragile edge cases, and the continuing need for expert validation. Particular attention will be given to AI memory management across long workflows, including how central modelling decisions, assumptions, and unresolved issues were documented to keep the process auditable.

In the Brightcon spirit, the session treats AI-assisted compilation as something that must be auditable rather than magical. It will argue that agentic AI can accelerate open sustainability data work, but only when embedded in reproducible workflows with explicit assumptions, decision logs, testable outputs, version control, and community review.