Beyond the LLM: Building Reliable AI Systems
Talk 1: Choosing the Right Intelligence for AI Workflows
Not every AI decision needs an LLM. Production systems often need to classify, score, route, generate, retry, wait, and recover, and different kinds of intelligence may be better suited to each task.
Using examples from the open-source Cadence AI Samples project, this talk explores how deterministic code, specialized decision models, LLMs, and agents can work together inside durable workflows. We’ll compare approaches through accuracy, decision consistency, latency, cost, and operational reliability, with the goal of choosing the simplest kind of intelligence that reliably solves each part of the system.
Talk 2: Building a Self-Evolving Mobile App
None of the existing consumer AI apps met my needs, so I built my own: OpenOx. In this talk, I’ll share how I did it so you can too! OpenOx is open source, and you can use it today. A useful agent needs to work with every service you use, from AI assistants like ChatGPT and everyday tools like Outlook to legacy government websites. But the world is too vast and dynamic for connectors built ahead of time. Instead, connectors can be created just in time for whatever service the user needs, which makes the app itself malleable: able to learn new interfaces, turn them into reusable capabilities, and evolve over time. Supporting any model, including local ones, lets it keep up with frontier intelligence without being locked into a single provider. Running on the user’s own device also makes web interactions more reliable: cloud agents often operate from centralized IPs that are easier for bot-protection systems to identify and block, while local agents use the user’s own IP and cookies. It also keeps more sensitive context, credentials, and browsing state local.
About Speakers
Kevin Burns is a Senior Developer Advocate on Uber's Cadence team, an OSS distributed workflow orchestration platform. His 25+ year software career also includes time at Microsoft and Volkswagen. Kevin co-founded Conversay and created the first commercially available voice browser, an early application of speaker-independent AI. At Volkswagen, he led developer onboarding for connected-vehicle APIs and built a GPT-powered support assistant grounded in platform documentation, vehicle logs, and internal system data. Today, his work increasingly focuses on the intersection of AI and durable workflow orchestration, including Cadence AI Samples, a project exploring how AI agents, models, and probabilistic decision-making can be combined with reliable, long-running workflows.
Ziyuan (Zi) Zhu is a software engineer at Amazon, where he has built some of the company’s most widely used internal agent-facing tools and products and leads an internal community focused on malleable software. Outside of work, Zi is an avid tennis player and created the Matcha Tennis app, a nonprofit passion project built to make tennis more accessible.
Thank you Uber for making this event possible.