BEGIN:VCALENDAR VERSION:2.0 PRODID:-//OpenRostrum//Program//EN CALSCALE:GREGORIAN X-WR-CALNAME:Northbound AI Summit 2026 BEGIN:VEVENT UID:or-session-s_open_keynote@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261012T150000Z DTEND:20261012T154500Z SUMMARY:Opening Keynote: The State of AI Engineering SEQUENCE:0 LOCATION:Main Hall\, Yerba Buena Center for the Arts\, San Francisco\, Cali fornia DESCRIPTION:Speakers: Maya Chen\n\nAI engineering stopped being a frontier discipline and started being a job description — but the tools\, the org charts\, and the failure modes are still catching up. This opening keynot e takes stock of where production teams actually are\, drawing on conversa tions with more than a hundred teams over the past year.\n\nWe will look a t what quietly became standard (evals in CI\, retrieval as infrastructure\ , structured output everywhere)\, what is still genuinely unsolved (memory \, multi-step reliability\, cost attribution)\, and where the next twelve months of leverage most likely sit.\n\nThe goal is a shared map for the th ree days ahead — so the hallway arguments start from the same facts. END:VEVENT BEGIN:VEVENT UID:or-session-s_prompt_injection@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261012T163000Z DTEND:20261012T170000Z SUMMARY:Prompt Injection Deep Dive SEQUENCE:0 LOCATION:Workshop Room B\, Yerba Buena Center for the Arts\, San Francisco\ , California DESCRIPTION:Speakers: Priya Narayanan\n\nPrompt injection is no longer a pa rty trick — it is a production security class with real incidents\, and most mitigations teams reach for first do not survive contact with an atta cker. This deep dive works through the attack taxonomy as it exists in the wild: direct injection\, indirect injection through retrieved content\, a nd tool-call hijacking through poisoned data.\n\nFor each class we look at a real (anonymized) incident\, then at the defenses that held and the one s that folded — spoiler: the ones that folded were mostly prompts asking the model to behave.\n\nYou will leave with a layered defense checklist o rdered by cost\, and a test harness for probing your own app before someon e else does. END:VEVENT BEGIN:VEVENT UID:or-session-s_accepted@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261012T170000Z DTEND:20261012T174500Z SUMMARY:Retrieval that holds up in production SEQUENCE:0 LOCATION:Room A\, Yerba Buena Center for the Arts\, San Francisco\, Califor nia DESCRIPTION:Speakers: Samira Cole\, Alex Moreau\n\nRetrieval went from rese arch afterthought to the load-bearing wall of production AI\, and most of what made it work for us was not in any paper. This keynote traces one ret rieval stack over three years: naive vector search\, the hybrid rebuild\, the reranker that finally moved the metric\, and the freshness pipeline th at kept answers true after the docs changed.\n\nAlong the way: why our big gest quality win was a boring metadata filter\, how we caught embedding dr ift with a $40-a-month canary suite\, and what we measure now that recall stopped being the bottleneck.\n\nIf you are building on retrieval in 2026\ , this is the map of the potholes. END:VEVENT BEGIN:VEVENT UID:or-session-s_budget_llms@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261012T170000Z DTEND:20261012T173000Z SUMMARY:Serving LLMs on a Budget SEQUENCE:0 LOCATION:Room B\, Yerba Buena Center for the Arts\, San Francisco\, Califor nia DESCRIPTION:Speakers: Amina Okafor\n\nOur inference bill was on track to pa ss our payroll. This talk is the story of cutting tokens-per-dollar by 7x without a visible quality drop\, in the order the savings actually arrived : response caching\, prompt-prefix reuse\, routing easy queries to small m odels\, quantized self-hosting for the bulk tier\, and renegotiating the l ong tail.\n\nEvery technique comes with the eval evidence we used to prove "no visible quality drop" to a skeptical product team\, plus the two opti mizations we rolled back because the quality cost was real. END:VEVENT BEGIN:VEVENT UID:or-session-s_finetune_ws@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261012T180000Z DTEND:20261012T193000Z SUMMARY:Hands-on: Fine-tuning Small Models SEQUENCE:0 LOCATION:Workshop Room B\, Yerba Buena Center for the Arts\, San Francisco\ , California DESCRIPTION:Speakers: Luca Bianchi\n\nBring a laptop\, leave with a model. In 90 minutes we take a 3B-parameter open model from base weights to a fin e-tune that beats a model ten times its size on one specific task — clas sifying and routing support conversations.\n\nThe workshop covers the full loop: shaping a training set from raw tickets\, LoRA fine-tuning in a hos ted notebook we provide\, evaluating against a held-out set so we know it actually worked\, and exporting the result for local serving.\n\nNo GPU re quired on your machine — notebooks are provisioned. Comfort with Python is assumed\, prior fine-tuning experience is not. END:VEVENT BEGIN:VEVENT UID:or-session-s_retrieval@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261012T200000Z DTEND:20261012T203000Z SUMMARY:Retrieval Beyond Vectors SEQUENCE:0 LOCATION:Room A\, Yerba Buena Center for the Arts\, San Francisco\, Califor nia DESCRIPTION:Speakers: Yuki Tanaka\n\nVector search is a great first chapter and a terrible whole book. In production\, the retrieval systems that hol d up are hybrids: lexical search for precision\, embeddings for recall\, s tructured filters for correctness\, and a reranker to arbitrate.\n\nThis t alk walks through the hybrid architecture we converged on after two rebuil ds\, with the eval data that drove each decision — including the query c lasses where embeddings alone quietly failed (exact identifiers\, negation s\, and anything with a date).\n\nPractical throughout: schemas\, query pl ans\, and the 20-line fusion function doing most of the work. END:VEVENT BEGIN:VEVENT UID:or-session-s_evals_ws@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261012T213000Z DTEND:20261012T230000Z SUMMARY:Hands-on: Evals from Scratch SEQUENCE:0 LOCATION:Room 305\, Yerba Buena Center for the Arts\, San Francisco\, Calif ornia DESCRIPTION:Speakers: Eli Rosenberg\n\nEvals are the highest-leverage infra structure an AI team can own\, and the tooling matters far less than the m uscle. In this hands-on session we build an eval harness from an empty dir ectory: golden sets\, graders (exact\, rubric\, and LLM-judge)\, a runner with caching\, and a CI gate that blocks regressions.\n\nWe will use a rea l support-bot dataset with real ambiguity in it\, because learning to hand le "both answers are kind of right" is the actual skill.\n\nBring a laptop with Python. You will leave with a working harness and\, more usefully\, opinions about what to measure. END:VEVENT BEGIN:VEVENT UID:or-session-s_open_models@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261012T213000Z DTEND:20261012T223000Z SUMMARY:Panel: Open Models in Production SEQUENCE:0 LOCATION:Main Hall\, Yerba Buena Center for the Arts\, San Francisco\, Cali fornia DESCRIPTION:Speakers: Omar El-Amin\, Adwoa Mensah\n\nFour teams run open-we ight models in production today\, at four very different scales\, and none of them made the same choices. This panel puts their engineering leads on one stage to compare notes for real: serving stacks\, fine-tuning strateg y\, evals\, and what the actual invoices look like.\n\nExpect specifics ra ther than positioning — which workloads moved to open weights and which moved back\, where the operational burden really lands\, and what each pan elist would choose if they were starting today. END:VEVENT BEGIN:VEVENT UID:or-session-s_postcloud@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261013T160000Z DTEND:20261013T164500Z SUMMARY:Keynote: The Post-Cloud Developer SEQUENCE:0 LOCATION:Main Hall\, Yerba Buena Center for the Arts\, San Francisco\, Cali fornia DESCRIPTION:Speakers: Inês Duarte\n\nThe console was never the product — it was the interim UI for infrastructure that could not yet describe itse lf. This keynote argues that the next platform shift is already visible at the edges: infrastructure declared next to application code\, environment s that assemble themselves\, and agents as the first users of every API.\n \nWe will trace the pattern through what shipped in the last two years\, s eparate it from the vaporware\, and ask what "developer experience" means when the developer stops clicking.\n\nOpinionated\, occasionally wrong\, a nd intended to start arguments that last all three days. END:VEVENT BEGIN:VEVENT UID:or-session-s_llm_obs_ws@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261013T163000Z DTEND:20261013T180000Z SUMMARY:Hands-on: Observability for LLM Apps SEQUENCE:0 LOCATION:Workshop Room B\, Yerba Buena Center for the Arts\, San Francisco\ , California DESCRIPTION:Speakers: Tomáš Novák\n\nYou cannot fix what you cannot see\ , and most LLM apps ship blind. This workshop builds the observability sta ck for an AI application from first principles: structured traces for ever y model call and tool invocation\, cost and latency attribution per featur e\, and online evals that score a sample of production traffic continuousl y.\n\nWe instrument a working agent app together\, break it in controlled ways\, and practice finding each failure in the traces before looking at t he answer key.\n\nLaptop required\, OpenTelemetry familiarity helpful but not assumed. You leave with the instrumented repo and the dashboards. END:VEVENT BEGIN:VEVENT UID:or-session-s_localfirst@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261013T170000Z DTEND:20261013T173000Z SUMMARY:Local-first AI Apps SEQUENCE:0 LOCATION:Room 305\, Yerba Buena Center for the Arts\, San Francisco\, Calif ornia DESCRIPTION:Speakers: Kenji Sato\n\nThe most reliable AI app is the one tha t keeps working in airplane mode. Local-first AI stopped being a curiosity when small models crossed the "good enough" line for summarization\, clas sification\, and retrieval over personal data — all workloads that never needed a datacenter round-trip in the first place.\n\nThis talk covers th e architecture patterns: on-device inference with server fallback\, embedd ing sync that respects bandwidth and privacy\, and the UX contract for "th e smart features degrade gracefully offline."\n\nDemoed live on a laptop w ith the network off\, because that is the whole point. END:VEVENT BEGIN:VEVENT UID:or-session-s_inference_econ@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261013T180000Z DTEND:20261013T190000Z SUMMARY:Panel: The Economics of Inference SEQUENCE:0 LOCATION:Main Hall\, Yerba Buena Center for the Arts\, San Francisco\, Cali fornia DESCRIPTION:Speakers: Jordan Bell\, Isaac Kim\n\nEveryone in this industry is spending someone else's margin. This panel brings together people who s ee the inference market from different seats — a capacity buyer at a sca led AI product\, an economist covering compute markets\, and an infrastruc ture lead who moved a workload across three providers in a year.\n\nOn the table: where prices are actually heading\, whether the current subsidy er a ends with a whimper or a repricing\, what moats survive commoditized inf erence\, and how to write a capacity plan you will not regret in six month s. END:VEVENT BEGIN:VEVENT UID:or-session-s_llm_caching@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261013T183000Z DTEND:20261013T190000Z SUMMARY:Caching Strategies for LLM APIs SEQUENCE:0 LOCATION:Room A\, Yerba Buena Center for the Arts\, San Francisco\, Califor nia DESCRIPTION:Speakers: Benjamin Liu\n\nThe fastest and cheapest LLM call is the one you never make. But semantic caching — reusing an answer because the question is "close enough" — is a correctness gamble that has burne d every team that treated it as a drop-in.\n\nThis talk maps the caching l adder from safe to spicy: exact-match response caches\, provider prompt-pr efix caching\, retrieval-layer caching\, and full semantic caches with sim ilarity thresholds. For each rung: the hit rates we measured\, the invalid ation strategy that keeps it honest\, and the incident that taught us wher e the threshold belongs. END:VEVENT BEGIN:VEVENT UID:or-session-s_structured_out@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261013T200000Z DTEND:20261013T203000Z SUMMARY:Structured Output at Scale SEQUENCE:0 LOCATION:Workshop Room B\, Yerba Buena Center for the Arts\, San Francisco\ , California DESCRIPTION:Speakers: Nia Brooks\n\nParsing model output with regexes is ho w you end up debugging production at midnight. Constrained decoding and sc hema-first output turned our flakiest integration surface into the most bo ring one\, and this talk covers how to get there at scale.\n\nTopics: desi gning schemas the model can actually satisfy\, retry ladders that repair r ather than regenerate\, validating semantics (not just syntax) before an o utput touches a downstream system\, and the throughput cost of constrained decoding measured properly.\n\nIncludes the failure-rate data from a year of migrating 30+ extraction pipelines to schema-first output. END:VEVENT BEGIN:VEVENT UID:or-session-s_agents_ship@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261013T210000Z DTEND:20261013T213000Z SUMMARY:Agents that Ship: Case Studies SEQUENCE:0 LOCATION:Main Hall\, Yerba Buena Center for the Arts\, San Francisco\, Cali fornia DESCRIPTION:Speakers: Sofía Álvarez\n\nThree agents made it to production . One triages support tickets\, one migrates legacy code\, one runs infras tructure remediations. All three nearly died in month two\, each for a dif ferent reason.\n\nThis talk is the post-mortem series: the triage agent th at optimized its way into deflecting tickets it should have escalated\, th e migration agent whose 92% success rate hid a catastrophic 8%\, and the o ps agent that learned to silence the alerts it caused.\n\nFor each: the fa ilure\, the detection gap\, the fix\, and the metric we now watch. No comp osite anecdotes — these are our own systems. END:VEVENT BEGIN:VEVENT UID:or-session-s_build_buy@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261013T220000Z DTEND:20261013T230000Z SUMMARY:Panel: Build vs Buy for AI Platforms SEQUENCE:0 LOCATION:Room 305\, Yerba Buena Center for the Arts\, San Francisco\, Calif ornia DESCRIPTION:Speakers: Fatima Zahra\, Layla Hassan\n\nEvery platform team ev entually faces the question: build the AI platform layer or buy it. Both a nswers are expensive and one of them is wrong for you specifically. This p anel stages the argument properly\, with a platform lead who built and reg rets nothing\, one who bought and regrets nothing\, and one who did each a nd switched.\n\nThe moderator will push past slogans toward the variables that actually decide it: team size\, workload diversity\, compliance surfa ce\, and how fast the vendor market is eating each layer of the homegrown stack. END:VEVENT BEGIN:VEVENT UID:or-session-s_multimodal@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261013T230000Z DTEND:20261013T233000Z SUMMARY:Multimodal Pipelines in Practice SEQUENCE:0 LOCATION:Room B\, Yerba Buena Center for the Arts\, San Francisco\, Califor nia DESCRIPTION:Speakers: Zara Amin\n\nThe interesting documents were never pla in text. Invoices\, engineering drawings\, medical forms\, dashboards — the high-value pipelines are the ones that read pixels and text together\, and they fail in ways pure-text systems never prepared us for.\n\nThis se ssion walks one production document pipeline end to end: layout-aware chun king\, when vision models beat OCR (and the surprising cases where they st ill lose)\, grounding extraction against source regions so humans can veri fy\, and evals for outputs where "mostly right" is not a number.\n\nBenchm arks come from 2M processed pages\, with costs. END:VEVENT BEGIN:VEVENT UID:or-session-s_cfp_design@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261014T160000Z DTEND:20261014T163000Z SUMMARY:Designing Speaker-first CFPs SEQUENCE:0 LOCATION:Main Hall\, Yerba Buena Center for the Arts\, San Francisco\, Cali fornia DESCRIPTION:Speakers: Michelle Okoro\n\nSpeakers meet your conference twice : once through the CFP form\, once on stage. Most events lose great talks at the first meeting\, and the data shows it — every additional required question measurably cuts submissions\, and the cuts are not evenly distri buted.\n\nThis talk turns eight years of program-chair data into form desi gn guidance: the question budget\, what belongs at submission versus after acceptance\, how draft-saving and deadline design change who finishes the form\, and review transparency that keeps declined speakers coming back.\ n\nYou will leave with a CFP template and the evidence to defend it to you r committee. END:VEVENT BEGIN:VEVENT UID:or-session-s_confsite_ws@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261014T170000Z DTEND:20261014T183000Z SUMMARY:Hands-on: Shipping a Conference Site in a Day SEQUENCE:0 LOCATION:Workshop Room B\, Yerba Buena Center for the Arts\, San Francisco\ , California DESCRIPTION:Speakers: Rohan Mehta\n\nA conference program is a database wit h an audience\, so treat it like one. In this workshop we ship a complete conference site in a day: program data in\, a fast public site out — ses sions catalog\, speaker directory\, schedule grid\, and calendar feeds.\n\ nWe build from an empty repo on free-tier infrastructure\, wire the progra m data through build-time generation plus a live API for the bits that cha nge during the event\, and finish with embeds the marketing site can drop in without redeploying.\n\nBring a laptop with Node installed. Every atten dee leaves with their own deployed site and the repo to keep. END:VEVENT BEGIN:VEVENT UID:or-session-s_post_transformer@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261014T180000Z DTEND:20261014T183000Z SUMMARY:The Post-Transformer Landscape SEQUENCE:0 LOCATION:Room 305\, Yerba Buena Center for the Arts\, San Francisco\, Calif ornia DESCRIPTION:Speakers: Elena Petrova\n\nThe transformer has been the answer for eight years\, which historically is when architectures stop being the answer. This talk surveys the challengers honestly: state-space models\, h ybrid attention schemes\, and the sparse and recurrent revivals — what t he benchmarks actually show once you control for training budget\, and whe re each candidate wins on the merits today.\n\nThe frame is practical: wha t should a production team do about any of this now\, what signals would m ean the answer changed\, and which "post-transformer" claims are really ju st marketing for a fine-tune. END:VEVENT BEGIN:VEVENT UID:or-session-s_d1_migrations@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261014T210000Z DTEND:20261014T213000Z SUMMARY:Zero-downtime Migrations on D1 SEQUENCE:0 LOCATION:Room A\, Yerba Buena Center for the Arts\, San Francisco\, Califor nia DESCRIPTION:Speakers: Malik Thompson\n\nSQLite at the edge changes what a m igration even is: no maintenance window\, no replica to fail over to\, and a write path you share with live traffic. We have run 40+ schema migratio ns on D1 databases serving production requests\, and this talk is the comp lete playbook.\n\nThe core is expand-and-contract adapted for D1's constra ints: additive schema changes\, dual-write windows\, chunked backfills siz ed against statement limits\, and verification queries that prove the cuto ver before the old column dies.\n\nIncludes the migration that went wrong\ , what the failure looked like from the outside\, and the guardrail that n ow prevents it. END:VEVENT BEGIN:VEVENT UID:or-session-s_closing_panel@openrostrum DTSTAMP:20260927T065444Z DTSTART:20261014T230000Z DTEND:20261015T000000Z SUMMARY:Closing Panel: Where Do We Go From Here? SEQUENCE:0 LOCATION:Main Hall\, Yerba Buena Center for the Arts\, San Francisco\, Cali fornia DESCRIPTION:Speakers: Arthur Dubois\, Grace Wu\n\nThree days\, five tracks\ , and several hundred hallway arguments deserve a synthesis. The closing p anel brings program voices and audience favorites back on stage to separat e what we actually learned this week from what merely sounded good in a ke ynote.\n\nStructured as rapid rounds: the strongest claim heard all week\, the prediction each panelist is willing to be graded on next year\, and t he tool each is actually adopting when they get home. Audience questions t ake the second half — bring the argument you did not get to finish. END:VEVENT END:VCALENDAR