BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.devconf.info//devconf-cz-2026//talk//GVQZTB
BEGIN:VTIMEZONE
TZID:CET
BEGIN:STANDARD
DTSTART:20001029T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20000326T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-devconf-cz-2026-GVQZTB@pretalx.devconf.info
DTSTART;TZID=CET:20260618T123000
DTEND;TZID=CET:20260618T130500
DESCRIPTION:You're paying GPT-4 prices for queries a smaller model could ha
 ndle. Model routing promises 40-70% savings.\n\nBut there's a catch: acade
 mic approaches assume humans label every response "good" or "bad." In prod
 uction\, those labels rarely exist.\n\nThis talk shows how to build a Baye
 sian router using Thompson Sampling that learns without labels. The key: a
  Composite Reward Function that scores responses automatically — Did the
  output parse? How fast was it? Did the agent retry? Three signals\, zero 
 human effort\, one score to update routing.\n\nWe also address production 
 realities often skipped in research:\n- Model rot: Adapting when a provide
 r degrades using decaying memory\n- Cold-start: Converging in 20 queries i
 nstead of 100 using expert priors\n- Safety: Continuous shadow evaluation 
 to guarantee accuracy\n\nResult: 40-50% cost cut\, <1% accuracy drop.\nPre
 requisites: Python\, LLM API familiarity.\n\n\nGitHub: https://github.com/
 shrinidhi-mahishi/model_routing\nPyPI: pip install bayesian-router
DTSTAMP:20260430T131558Z
LOCATION:D0207 (capacity 90)
SUMMARY:The 50% Cheaper Agent: Autonomous LLM Routing with Bayesian Bandits
  - Shrinidhi Mahishi
URL:https://pretalx.devconf.info/devconf-cz-2026/talk/GVQZTB/
END:VEVENT
END:VCALENDAR