BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Session Canvas//Public Embed//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:DevFlow Conf 2027
BEGIN:VEVENT
UID:cmsky602w008vn7pksa7j1xc1@sessioncanvas
DTSTAMP:20260811T000935Z
DTSTART:20270512T160000Z
DTEND:20270512T164500Z
SUMMARY:Shipping Agents That Don't Melt Down at 3 A.M.
LOCATION:Golden Gate Ballroom
DESCRIPTION:Speakers: Samuel Adeyemi\n\nA field report on running long-live
 d agents in production: timeouts\, retries\, poison inputs\, and the runbo
 ok that finally made the pager quiet.\n\nWe started with one agent that co
 uld file a support ticket and ended up with a fleet of them touching billi
 ng\, inventory and customer email. Everything that broke\, broke at night 
 — a tool call that hung for nineteen minutes\, a retry loop that re-sent
  the same refund four times\, and a prompt change that quietly doubled tok
 en spend on the busiest queue of the week.\n\nThis keynote walks through t
 he three incidents that reshaped our architecture and the guardrails we sh
 ipped after each one: hard per-step deadlines\, idempotency keys on every 
 side effect\, a spend ceiling enforced outside the model\, and a replay vi
 ew that lets an on-call engineer see exactly what the agent believed at th
 e moment it went wrong.\n\nYou will leave with the runbook we now hand to 
 every team before their agent gets production credentials\, the four dashb
 oards we page on\, and an honest account of which safeguards were worth th
 e engineering time and which ones we deleted six weeks later.
STATUS:CONFIRMED
END:VEVENT
BEGIN:VEVENT
UID:cmsky60b0008xn7pkyfowpm74@sessioncanvas
DTSTAMP:20260811T000935Z
DTSTART:20270512T171500Z
DTEND:20270512T174000Z
SUMMARY:Evals as a Product Surface: Making Quality Legible
LOCATION:Mission Hall
DESCRIPTION:Speakers: Diego Ferreira\n\nEvaluations are usually treated as 
 an internal chore. We shipped ours to customers instead\, and it changed h
 ow the whole team argued about quality.
STATUS:CONFIRMED
END:VEVENT
BEGIN:VEVENT
UID:cmsky60gs008zn7pkj7lapja9@sessioncanvas
DTSTAMP:20260811T000935Z
DTSTART:20270512T180000Z
DTEND:20270512T182500Z
SUMMARY:The RAG Is Dead\, Long Live Retrieval
LOCATION:Presidio Room
DESCRIPTION:Speakers: Amara Okonkwo\n\nThree years of retrieval architectur
 es\, what survived\, what did not\, and why the boring parts (chunking\, f
 reshness\, permissions) still decide the outcome.
STATUS:CONFIRMED
END:VEVENT
BEGIN:VEVENT
UID:cmsky6210009in7pk1kjy1fi0@sessioncanvas
DTSTAMP:20260811T000935Z
DTSTART:20270512T193000Z
DTEND:20270512T195500Z
SUMMARY:Sponsor Showcase: Kestrel Compute
LOCATION:Golden Gate Ballroom
DESCRIPTION:Speakers: Jonas Weber\n\nA twenty-minute look at how Kestrel Co
 mpute schedules mixed CPU/GPU workloads\, presented by our platinum sponso
 r.
STATUS:CONFIRMED
END:VEVENT
BEGIN:VEVENT
UID:cmsky60ok0091n7pklyeqd975@sessioncanvas
DTSTAMP:20260811T000935Z
DTSTART:20270512T203000Z
DTEND:20270512T220000Z
SUMMARY:Cutting Inference Costs 70% Without Touching the Model
LOCATION:Workshop Loft
DESCRIPTION:Speakers: Jonas Weber\, Priya Ramanathan\n\nCaching\, batching\
 , routing and quantisation\, measured end to end on a real workload. Inclu
 des the changes that looked clever and made things worse.\n\nOur inference
  bill grew faster than our traffic for two quarters straight\, and the obv
 ious answers — a smaller model\, a cheaper provider — were off the tab
 le for quality reasons. So we went looking for the money everywhere else: 
 in the cache we were not reusing\, in the requests we were sending one at 
 a time\, and in the long tail of prompts nobody had read since launch.\n\n
 In this hands-on session we rebuild that programme from scratch against a 
 live workload. You will add a prompt cache and measure the real hit rate r
 ather than the hoped-for one\, batch a stream of requests and watch the la
 tency percentiles you just traded away\, route by difficulty so the expens
 ive path only handles the requests that need it\, and trim context that ha
 s been dead weight since the second sprint.\n\nWe finish with the cost mod
 el we now review monthly — cost per resolved request rather than cost pe
 r token — plus the two optimisations we rolled back and the graphs that 
 told us to. Bring a laptop\; every exercise ships with a workload you can 
 run locally.
STATUS:CONFIRMED
END:VEVENT
BEGIN:VEVENT
UID:cmsky60ue0094n7pkmu47qo1q@sessioncanvas
DTSTAMP:20260811T000935Z
DTSTART:20270513T163000Z
DTEND:20270513T165500Z
SUMMARY:Structured Output Is a Distributed Systems Problem
LOCATION:Golden Gate Ballroom
DESCRIPTION:Speakers: Mei-Lin Chao\n\nSchema drift\, partial responses and 
 retries: why validating model output belongs in the same mental model as a
 ny other unreliable network call.
STATUS:CONFIRMED
END:VEVENT
BEGIN:VEVENT
UID:cmsky61260096n7pkhsi1scsa@sessioncanvas
DTSTAMP:20260811T000935Z
DTSTART:20270513T180000Z
DTEND:20270513T193000Z
SUMMARY:Building a Tool-Calling Runtime in 300 Lines
LOCATION:Embarcadero Studio
DESCRIPTION:Speakers: Tobias Andersson\n\nLive-coded from an empty file: a 
 minimal\, debuggable tool-calling loop with cancellation\, tracing and a s
 andbox\, and what each abstraction buys you.
STATUS:CONFIRMED
END:VEVENT
BEGIN:VEVENT
UID:cmsky618n0098n7pkih8yxv9z@sessioncanvas
DTSTAMP:20260811T000935Z
DTSTART:20270514T170000Z
DTEND:20270514T172500Z
SUMMARY:Observability for Nondeterministic Systems
LOCATION:Mission Hall
DESCRIPTION:Speakers: Yuki Tanaka\n\nTraces\, spans and sampling when the s
 ame input legitimately produces different output. What to log\, what to re
 dact\, and how to make it searchable.
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR