| |
|
|
Building AI Applications At Scale
|
| With Robert Hodges (SVP for Real-Time Analytics, EnterpriseDB), Matt Wimpelberg (Senior Solutions Architect, Grafana Labs), Brad Miro (Senior Developer Advocate, Google Cloud). |
|
Dec 08 (Tue) @ 06:00 PM
FREE
| | 159 W 25th St, 159 W 25th St, 3rd Fl
|
|
|
|
|
| |
| Register |
|
|
| |
AI demos are everywhere.
But how do companies actually build AI products that work reliably with live data, low latency, & thousands of users?
Join engineers as they share the architectures behind production AI systems - from streaming data & event-driven architectures to real-time databases, RAG pipelines, observability, & open-source infrastructure.
No hype. Just practical engineering stories, lessons learned, & opportunities to meet others building modern data systems.
Food & drinks included
Agenda
6 pm - Networking
6:15 - 8:00 pm - Talks
8:00 - 9:00 pm - Networking
Speakers
Robert Hodges, SVP for Real-Time Analytics @ EnterpriseDB
Matt Wimpelberg, Senior Solutions Architect @ Grafana Labs
Brad Miro, Senior Developer Advocate @ Google Cloud
____
Description of the talks
Tools & Tricks to Prepare Your ClickHouse Clusters for AI
Speaker: Robert Hodges, SVP for Real-Time Analytics @ EnterpriseDB
Abstract: AI is coming! It's going to crush your ClickHouse database!! Well maybe. This talk will cut through the FUD, focusing on three ground-level issues that are emerging across a wide range of installations. First, build sandboxes to protect data if (when) your agents go rogue. Second, lower costs of operating ClickHouse to free up resources for AI. Third, integrate AI agents into your development & administrative workflows. We're working on all three at EDB. We'll share learnings as well as our long-term roadmap for AI & open source databases.
Observing LLMs Without the Hype: How far boring tools get you, & what broke anyway
Speaker: Matt Wimpelberg, Senior Solutions Architect @ Grafana Labs
Abstract: Most teams shipping LLM features already run Prometheus, a trace store & a log pipeline. This talk asks how far that stack gets you before you need anything LLM-specific.
We start with what OpenTelemetry has actually standardized for generative AI (spans, two client metrics, token & model attributes) & what is still experimental or missing. Then, in short recorded demos, we run a plain open source stack against a simulated LLM service & watch three assumptions break:
- Cardinality: user & conversation IDs on metrics.
- Sampling: a 5% head sampler discards the errors & the slow tail you care about.
- Payload size: prompts & completions are tens of kilobytes, not tens of bytes.
For each, we show the boring fix & what the fix costs.
We close with a decision rule for when a dedicated LLM platform on a columnar store is worth adding. That is when you need to query prompt & completion content at volume, join it with evals & feedback, or slice cost by high-cardinality dimensions.
Attendees will leave with
1. A map of what the OpenTelemetry GenAI conventions cover today, & where they stop.
2. Three concrete failure modes of LLM telemetry in a classic metrics, traces & logs stack, with the config that fixes each.
3. A short checklist for deciding whether they have outgrown that stack.
Other talks coming soon!
Follow the OSA Community
Make OSA Con your community!
Website: www.osacom.io
LinkedIn: /osa-con
Slack: Join our community
|
|
|
|
|
|
|
|