I built this partly because i wanted a cleaner side-by-side comparison of the big “2026 outlook” reports (banks + asset managers), and partly as an experiment in whether codex / claude code / cursor together can be pushed to handle citations accurately rather than hallucinating summaries. Quite pleased with the result.
The workflow was roughly: ingest publicly available outlook reports, extract claims, numbers, and framing (& force every synthesized statement to be tied back to a specific source/page), anchor on key themes (ai capex, policy path, power/energy, credit, etc.) rather than by institution.
The workflow was roughly: ingest publicly available outlook reports, extract claims, numbers, and framing (& force every synthesized statement to be tied back to a specific source/page), anchor on key themes (ai capex, policy path, power/energy, credit, etc.) rather than by institution.