Jinwoo Je | Portfolio

Amazon Last Mile's forecasting system, UPF, is launching with the engine (PRIMA) and product (Armada) that I built. PRIMA outperformed the Research Science team's model in parallel shadow test WAPE, plan-over- plan volatility, metric explainability, and execution time. Armada provides better workflow integration than the existing software suite. This wasn't intended—PRIMA was built as a baseline to beat. This decision required focusing on production readiness: HVE-specific modeling, region expansion assumptions, and redirecting resources from RS model support to PRIMA while building input pipelines with partner tech teams.
I'm revising my views on multitasking and context switching. This contradicts Deep Work, but with thinking models, I'm no longer doing the work directly. I collaborate with agents to create and align on plans, hand off for iteration, then check back in. Getting comfortable managing parallel work streams— multitasking 2.0—might become legitimate.
Side project progress has been slower than expected and holding it on pause to fully focus on UPF. Project complexity increases with deeper understanding and better tools. I've seen significant AI benefits in familiar work, but in new areas, "not knowing what I don't know" becomes a blocker when stuck. Detecting architectural issues versus knowing when to push through remains challenging. Learning the domain and specific technology to some degree is essential for quality work.