Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Defensive LLM Calling
Search
Sponsored
·
Your Podcast. Everywhere. Effortlessly.
Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
→
Ogata Katsuya
December 05, 2025
0
170
Defensive LLM Calling
Ogata Katsuya
December 05, 2025
Tweet
Share
More Decks by Ogata Katsuya
See All by Ogata Katsuya
AIで加速する文化祭ソフトウェア開発 / Software Development for School Festivals Accelerated by AI
ogatakatsuya
0
39
Vibe-Coding で作ったアプリを 文化祭で運用した話 / The story of running an app I built with Vibe-Coding at a school festival
ogatakatsuya
1
160
ハッカソンの勘所とエンジニアリングへの活かし方 / What Hackathons Teach Us and How to Bring That into Engineering
ogatakatsuya
0
41
Kubernetesで分散処理をやってみる-Kafkaを添えて-
ogatakatsuya
0
230
ソフトウェアアーキテクチャ入門 / Introduction to Software Architecture
ogatakatsuya
2
150
大学のサークルプラットフォームを作った話
ogatakatsuya
0
120
Go College
ogatakatsuya
0
140
文化祭で使うアプリを1人で作った話
ogatakatsuya
0
110
twitter-cloneを作った話
ogatakatsuya
0
66
Featured
See All Featured
The Web Performance Landscape in 2024 [PerfNow 2024]
tammyeverts
12
1k
Ten Tips & Tricks for a 🌱 transition
stuffmc
0
82
Six Lessons from altMBA
skipperchong
29
4.2k
It's Worth the Effort
3n
188
29k
Designing for Timeless Needs
cassininazir
0
150
Building Applications with DynamoDB
mza
96
6.9k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
34
2.6k
Save Time (by Creating Custom Rails Generators)
garrettdimon
PRO
32
2.3k
Ethics towards AI in product and experience design
skipperchong
2
210
Bash Introduction
62gerente
615
210k
Claude Code のすすめ
schroneko
67
220k
End of SEO as We Know It (SMX Advanced Version)
ipullrank
3
4k
Transcript
Kyoto Tech Talk #9 2025/12/05 Defensive LLM Calling Katsuya Ogata
ࣗݾհ • ໊લ: ॹํ ࠀ࠸ʢ͓͕ͨ ͔ͭʣ • ॴଐ: େࡕେֶ ใՊֶݚڀՊ
• झຯ: ొࢁɾཱྀߦɾαφ • MLLMΛ༻͍ͨಈըղੳͷݚڀ • ϝϧϘϧϯʹߦͬͯ·ͨ͠ • X: @ogata_katsuya • ݸਓ: www.ogatakatsuya.com ϝϧϘϧϯཱྀߦ ޒ৭ϲݪͷొࢁ
None
AWS 2025/10/20 Google Cloud 2025/06/13 Azure 2025/10/29
ΫϥυյΕΔલఏͰར༻͢Δ
2025ʹো͕ى͖ͯͨͷʁ
શ෦
োͷݪҼͱ༰ • ݪҼ • AIΛϗετ͍ͯ͠ΔΫϥυͷো • ίϯύΠϥͷόάʢΠϯϑϥىҼʣ • ༰ •
ϨΠςϯγͷٸܹͳ૿Ճ • ਫ਼ͷԼ • ͦͦLLMΛݺͼग़͢͜ͱ͕Ͱ͖ͳ͍
LLMյΕΔલఏͰར༻͢Δ
🤔
Application Layer Transport Layer Network Layer Network Access Layer TCP/IP
τϥϯεϙʔτ • ΞυϨεϙʔτ൪߸ͷΈʢIPΛͬͯΔͷωοτϫʔΫʣ • ͲΜͳωοτϫʔΫͷܦ࿏Λ௨Δ͔Βͳ͍ • ωοτϫʔΫͦͷͷΛ৴པ͍ͯ͠ͳ͍લఏͰ࣮֬ʹσʔλΛૹΓಧ͚Δ • ࠶ૹ੍ޚɺ੍ޚɺޡΓݕग़ •
ʢͷߟ͑Δʣτϥϯεϙʔτͷׂ • ΞϓϦέʔγϣϯʹωοτϫʔΫͷΓޱΛݟͤΔ • ΞϓϦέʔγϣϯωοτϫʔΫͷҧ͍Λҙࣝ͠ͳͯ͘ྑ͍ • ωοτϫʔΫͦͷͷͷ৴པੑΛ୲อ͢Δ
🤔
Application Layer Transport Layer Network Layer Network Access Layer TCP/IP
Application Endpoint ????? LLM Calling LLM Calling ͜͜
LLM Callingʹ͓͚Δτϥϯεϙʔτ • LLM Providerͷҧ͍Λٵऩͯ͠ɺͦͷΓޱ͚ͩΛݟͤΔ • நԽ • Ӆṭ •
LLM ProviderΛ৴པ͍ͯ͠ͳ͍લఏͰ࣭Λ୲อ͢Δ • ϦτϥΠॲཧ • ଞͷProviderͷϑΥʔϧόοΫ • ඞͣԿ͔͠ΒͷϨεϙϯε͕ฦͬͯ͘Δ
PythonͰͷ࣮ྫ abcΛ༻͍ͨநΫϥεͷ࡞ LiteLLMͱ͍͏ศརϥΠϒϥϦ͋Γ·͢
PythonͰͷ࣮ྫ ֤ϓϩόΠμʔͰͷநϝιουͷ࣮
PythonͰͷ࣮ྫ LLM Callingͷτϥϯεϙʔτ(͔ͳΓંͬͯ·͢)
LLM Callerઢ • τϥϯεϙʔτʹඞཁͳύϥϝʔλΛ͚ͩ͢ • ͲΜͳϓϩόΠμͰ࣮ߦ͞ΕΔ͔Βͳ͍ • ΠϯλʔϑΣʔε͚ͩΛҙࣝͯ͠ɺͦΕҎԼΛؾʹ͠ͳ͍ • ͚ͩͲɺҰఆ࣭͕อͨΕͯɺ͋ΔఔͷͰฦ͕ฦͬͯ͘Δ
ສࣄղܾʁʁʁ • LLMͷোૉʹΤϥʔΛు͍ͯམͪͯ͘Εͳ͍ • Ϩεϙϯε͕500൪ͳΒେৎ • ϨΠςϯγ͕૿Ճ͢Δ • λΠϚʔͱϦτϥΠॲཧͰͳΜͱ͔ͳΔʢ͔ʣ •
ਫ਼͕མͪΔ • োΛײ͢Δ͜ͱ͕͍͠ • ࠷ऴతʹɺ͜·ΊͳࢹͱɺϢʔβʔͷʹৗʹහײͰ͋Δ͜ͱ͕ॏཁ • LangfuseLangsmithͳͲͷLLMࢹମ੍ͷඋඞཁ
·ͱΊ • ͜ͷLTʹ໌֬ͳ͑͋Γ·ͤΜ • LLMͷ࣭Λ୲อ͢ΔͷϓϩόΠμͪΖΜɺϢʔβʔΛෛ͍ͬͯ Δͱ͍͏ఏى • ࠷ݶఏڙ͖͢αʔϏεͷ࣭Λ୲อ͢ΔͨΊͷϑΥʔϧόοΫઓུΛߟ͑ ͍ͯ͘ඞཁ͕͋Δ •
օ͞Μ͕࣮ફ͍ͯ͠ΔϑΥʔϧόοΫઓུɾख๏Λڭ͍͑ͯͩ͘͞
Reference • https://azure.status.microsoft/en-us/status/history/ • https://www.youtube.com/watch?v=D297rfqSFvY • https://status.cloud.google.com/incidents/ow5i3PPK96RduMcb1SsW • https://www.mobihealthnews.com/news/google-cloud-outage-disrupts-ai- services-used-healthcare
• https://www.youtube.com/watch?v=Ar_ckPSwe1Y • https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues