BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//pydata-amsterdam2026//speaker//JXHB9D
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:Rebuilding Picnic’s Recipe Recommender - Thijs Sluijter\, Majid 
 Hajiheidari
DTSTART;TZID=Europe/Amsterdam:20260911T141000
DTEND;TZID=Europe/Amsterdam:20260911T144000
DTSTAMP:20260911T153631Z
UID:pretalx-pydata-amsterdam2026-VUJB7A@pretalx.com
DESCRIPTION:Over the past year\, we rebuilt our recipe recommender at Picn
 ic. The project was prompted by a new recipe application UI\, which change
 d our recommendation problem from retrieval to ranking. This forced us to 
 rethink our recommendation pipeline by modularising it to allow personalis
 ation in any context across our app\, and rethinking our model choice from
  a fast\, retrieval-optimised two-tower model to an expressive Deep & Cros
 s Network (DCN) ranking model. Running these experiments at scale required
  rethinking the data pipeline\, too. We rebuilt our training data pipeline
  using Polars and Apache Arrow to process tens of millions of interactions
  efficiently\, construct purchase histories with lower memory overhead\, a
 nd move features into PyTorch with minimal copying. Finally\, we built a c
 onstrained Autoresearch loop in which coding agents proposed configuration
 s\, ran training\, evaluated results\, and recorded the next hypothesis\, 
 all while data splits\, metrics\, production constraints\, and reproducibi
 lity controls remained fixed.
LOCATION:The Grid
URL:https://pretalx.com/pydata-amsterdam2026/talk/VUJB7A/
END:VEVENT
END:VCALENDAR