BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.coscup.org//coscup-2026//talk//NJSSSE
BEGIN:VTIMEZONE
TZID:CST
BEGIN:STANDARD
DTSTART:20000101T000000
RRULE:FREQ=YEARLY;BYMONTH=1
TZNAME:CST
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-coscup-2026-NJSSSE@pretalx.coscup.org
DTSTART;TZID=CST:20260809T114500
DTEND;TZID=CST:20260809T121500
DESCRIPTION:Most applications treat AI as a cloud-only feature — send a r
 equest\, wait for a response\, pay per token. Open models like Gemma\, Lla
 ma\, DeepSeek\, and Phi freed developers from proprietary APIs\, but not f
 rom costs: you still need to host them somewhere — GPU servers\, infrast
 ructure\, per-request pricing. You can run models directly on the user's d
 evice — zero inference costs\, offline access\, full data privacy. But t
 here's a trade-off: models small enough to run on a phone or in a browser 
 can't handle every task.\nHybrid AI is the answer to this dilemma. Combine
  cloud and on-device models in a single application: simple and privacy-se
 nsitive tasks run on-device\, complex reasoning and multimodal workloads g
 o to the cloud. The problem is that you end up maintaining two completely 
 different inference stacks: different runtimes (TFLite\, LiteRT-LM\, llama
 .cpp)\, different model formats (TFLite flatbuffers vs GGUF)\, different q
 uantization strategies\, and different streaming APIs — all within one a
 pp.\nIn this talk\, I'll show how to solve this problem with Genkit — an
  open-source AI framework from Google with an open plugin system. Each plu
 gin adapts a specific runtime or model format to a unified pipeline: genki
 t_flutter_gemma for TFLite/LiteRT models\, genkit_llamadart for GGUF. Engi
 ne-specific details are hidden behind a single API for flows\, structured 
 output\, tool calling\, and agentic workflows. Need a new runtime? Write a
  plugin. New model format? Same story. On-device inference runs across And
 roid\, iOS\, macOS\, Windows\, Linux\, and Web — switching between cloud
  and local execution is a one-line model reference change.\nWe'll explore 
 hybrid patterns in practice: how to route between on-device and cloud base
 d on model capabilities\, connectivity\, and task complexity. We'll cover 
 the real trade-offs — quantization impact on output quality\, memory con
 straints across platforms\, cold start latency\, and when hybrid actually 
 makes sense versus pure cloud or pure edge.\nOpen-source models made on-de
 vice AI possible. Open-source tooling makes hybrid AI extensible.
DTSTAMP:20260713T142152Z
LOCATION:AU
SUMMARY:Hybrid AI: The Next Pattern for AI-Powered Apps - Sasha Denisov
URL:https://pretalx.coscup.org/coscup-2026/talk/NJSSSE/
END:VEVENT
END:VCALENDAR