Show Notes
Why do most machine learning projects never leave the lab? The position here is blunt: the model is about ten percent of the work and the other ninety percent is pipelines, monitoring and ownership. The episode walks through the three failure points at the notebook to production handoff, data drift, unowned pipelines after launch, and systems never designed for retraining, and why a drifting model costs money quietly instead of paging you at midnight.
You come away able to name the roles that fix this, the ML engineer and the data engineer, read vendor claims from DBT Labs against O'Reilly Radar and MIT Sloan Management Review, apply data as a product, and ask the accountability question before funding an AI project.
0:00 Why ML projects die after the notebook 1:10 Three places the handoff snaps 2:02 The ML engineer nobody budgets for 2:25 Data as a product, not byproduct 3:25 Why buying a platform does not help 4:07 The one question to ask before greenlighting
More from MBA Training Data
Who pays when a fund's NAV is wrong and the manager never saw the error? Under CSSF Circular 24/856 the investment fund manager corrects and compensates, even when the administrator calculated the price. The episode argues that managers should own one position and pricing record, backed by a written pricing policy, rather than inheriting whatever the administrator used.
How did hackers send a ransom note through the real Asos app? Reporting points to a conned employee and, per researcher Kevin Beaumont, Simon AI, a composable customer data platform owned by Monetate that reads straight from Snowflake with MFA optional and off by default. That route is unconfirmed. The episode argues the real failure is ownership: marketing bought it, data engineering wired it to the warehouse, and nobody owned the login.
Agents dominated dbt Summit, but does smart orchestration or a context layer make data agent-ready on its own? The episode argues no. Fanatics Betting and Gaming saw about 0.2% model reuse from dbt State until it declared source freshness and model-level SLAs, which raised reuse to roughly 15%. The same pattern runs through Fivetran's Context Layer, AI audit trails and token costs.
How did an attacker using ARTEX, an open-source agentic AI penetration-testing tool, take data on more than 67,000 customers from seven South Korean firms, including Shinhan, KB Kookmin and Hana, without touching core banking? The episode argues the failure was data access design. Broker portals and employee apps returned income, loan limits and resident registration numbers to callers who did not need them, despite nearly 124 billion won in security spending.
What happens when a data platform is legally defensible, operationally embedded and publicly distrusted at the same time? More than 44,000 people have filed Article 21 objections under UK GDPR against NHS England's Palantir-powered Federated Data Platform, live in 142 trusts, weeks before a February 2027 contract break clause comes into reach. The position here: objection handling is a throughput problem, not a policy document.