Northbound AI Summit 2026

Bring a laptop, leave with a model. In 90 minutes we take a 3B-parameter open model from base weights to a fine-tune that beats a model ten times its size on one specific task — classifying and routing support conversations. The workshop c…

No slides, just terminals. This is a live tour of a production agent swarm doing real work: a fleet of specialized agents coordinating on tasks with checkpoints, retries, and a shared control plane — running against live infrastructure on s…

Vector search is a great first chapter and a terrible whole book. In production, the retrieval systems that hold up are hybrids: lexical search for precision, embeddings for recall, structured filters for correctness, and a reranker to arbi…

Prompt injection is no longer a party trick — it is a production security class with real incidents, and most mitigations teams reach for first do not survive contact with an attacker. This deep dive works through the attack taxonomy as it…

Retrieval went from research 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 retrieval stack over three years: naive vector search, the hybrid rebuil…

Evals are the highest-leverage infrastructure an AI team can own, and the tooling matters far less than the muscle. In this hands-on session we build an eval harness from an empty directory: golden sets, graders (exact, rubric, and LLM-judg…

Four teams run open-weight 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 strate…

Our inference bill was on track to pass 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 qu…

The console was never the product — it was the interim UI for infrastructure that could not yet describe itself. This keynote argues that the next platform shift is already visible at the edges: infrastructure declared next to application c…

You cannot fix what you cannot see, and most LLM apps ship blind. This workshop builds the observability stack for an AI application from first principles: structured traces for every model call and tool invocation, cost and latency attribu…

The most reliable AI app is the one that keeps working in airplane mode. Local-first AI stopped being a curiosity when small models crossed the "good enough" line for summarization, classification, and retrieval over personal data — all wor…

The 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 burned every team that treated it as a drop-in. This talk maps…

Everyone in this industry is spending someone else's margin. This panel brings together people who see the inference market from different seats — a capacity buyer at a scaled AI product, an economist covering compute markets, and an infras…

Parsing model output with regexes is how you end up debugging production at midnight. Constrained decoding and schema-first output turned our flakiest integration surface into the most boring one, and this talk covers how to get there at sc…

Three agents made it to production. One triages support tickets, one migrates legacy code, one runs infrastructure remediations. All three nearly died in month two, each for a different reason. This talk is the post-mortem series: the tria…

Every platform team eventually faces the question: build the AI platform layer or buy it. Both answers are expensive and one of them is wrong for you specifically. This panel stages the argument properly, with a platform lead who built and…

The interesting documents were never plain 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…

Speakers 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 e…

A conference program is a database with 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 — sessions catalog, speaker directory, schedule grid, and calen…

The 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, hybrid attention schemes, and the sparse and recurrent revi…

SQLite at the edge changes what a migration 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 migrations on D1 databases serving production requests, and this ta…

Three days, five tracks, and several hundred hallway arguments deserve a synthesis. The closing panel brings program voices and audience favorites back on stage to separate what we actually learned this week from what merely sounded good in…

Shipping an agent inside a 40,000-person company is a different sport from shipping one to developers. The blockers are rarely technical: procurement wants a vendor risk review, security wants an audit trail for every tool call, and legal w…

The subscription wall was a billing model that grew into an architecture, and it is quietly coming apart. Teams are replacing rented software with open cores they can read, run, and modify — not for ideology, but because the economics final…