Demand-Driven Context: What 300 Incidents Reveal About Tribal Knowledge - Raj Navakoti

This talk was recorded at NDC Oslo in Oslo, Norway. #ndcoslo #ndcconferences #developer #softwaredeveloper Attend the next NDC conference near you: https://ndcconferences.com https://ndcoslo.com/ Subscribe to our YouTube channel and learn every day: / @NDC Follow our Social Media! https://www.facebook.com/ndcconferences https://twitter.com/NDC_Conferences https://www.instagram.com/ndc_conferences/ #genai #ai #architecture #tdd RAG can't retrieve what was never documented. Fine-tuning can't learn what only exists as institutional habit. And yet that's exactly where enterprise knowledge lives — in chat threads, runbooks nobody reads, and the heads of experienced staff. We gave an AI agent 300 real production incidents from a retail e-commerce platform and asked it to resolve them. It failed — repeatedly, predictably, and in ways that told us exactly what knowledge was missing. It didn't know that the nightly batch job retries three times before escalating to the on-call team. That "fulfilment window expired" means something different in the Nordic region than in Central Europe. That the inventory service was migrated eighteen months ago but three teams still call it by the old name. We developed Demand-Driven Context (DDC), a methodology inspired by Test-Driven Development: instead of documenting everything upfront and hoping it's useful, we let the agent fail on real problems, identify exactly what knowledge is missing, and curate only the minimum context needed to succeed. Failure is the signal for what to document. After nine DDC cycles, the agent produced a reusable knowledge base of 46 entities — and we observed convergence: each new incident required fewer new entities because previous cycles had already captured overlapping knowledge. This talk walks through the methodology, the convergence data, and a scaling architecture for running DDC across domains.