BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//pydata-amsterdam2026//speaker//WVHGVT
BEGIN:VTIMEZONE
TZID:Europe/Amsterdam
BEGIN:DAYLIGHT
DTSTART:20250911T000000
TZNAME:CEST
TZOFFSETFROM:+0200
TZOFFSETTO:+0200
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20251026T030000
RDATE:20261025T030000
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20260329T030000
RDATE:20270328T030000
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
SUMMARY:Systems for Scale: Architecting a Nationwide Energy Forecasting Pl
 atform - Mohit Kumar\, Daria Mustafina\, Corné Vriends
DTSTART;TZID=Europe/Amsterdam:20260911T132500
DTEND;TZID=Europe/Amsterdam:20260911T135500
DTSTAMP:20260911T153628Z
UID:pretalx-pydata-amsterdam2026-CBE9UC@pretalx.com
DESCRIPTION:Standard MLOps stacks assume you are operating tens or hundred
 s of models. What happens when you need to train\, validate\, and serve th
 ousands of forecasting models on a weekly cadence\, producing millions of 
 forecasts for a national energy portfolio? Tools like MLflow start to buck
 le\, per-model evaluation stops being meaningful\, and the choice between 
 one global model and many local ones becomes an architectural decision rat
 her than a modeling one.\nThis talk is a systems architecture case study o
 f the platform we built when off-the-shelf MLOps stopped working. We focus
  on three concrete\, transferable patterns:\n\n1. Heterogeneous parallel t
 raining — running Spark ML global models and thousands of local models (
 via applyInPandas) in a single unified workflow\, with a four-axis decisio
 n framework (data volume per entity\, signal heterogeneity\, cold-start be
 havior\, operational cost) for routing each segment.\n2. A model registry 
 that scales beyond MLflow — what we kept\, what we dropped\, and the sch
 ema that lets us manage versions\, parameters\, and artifacts for a vast p
 ortfolio without metadata-store collapse.\n3. Portfolio-level promotion ga
 tes — why per-model accuracy metrics mislead at scale\, and how we run c
 hampion/challenger experiments with strict temporal integrity and aggregat
 e decision criteria.\n\nAttendees will leave with patterns they can apply 
 to any large-scale time-series system\, not just energy forecasting.
LOCATION:Entropy
URL:https://pretalx.com/pydata-amsterdam2026/talk/CBE9UC/
END:VEVENT
END:VCALENDAR