Program & speakers

Natural-language interfaces to databases promise self-serve analytics, but production Text-to-SQL agents fail loudly when schemas drift, questions are ambiguous, or SQL must be safe and auditable. This session focuses on engineering patterns that make Text-to-SQL agents more reliable: grounding in schema and business semantics, validation and guardrails before execution, observability for bad generations, and iterative evaluation so teams can ship with confidence on AWS.

Artificial Intelligence (AI) stands at a pivotal point in global technological evolution, simultaneously perceived as a threat, a tool, and a transformation partner. As a threat, AI raises legitimate concerns around automation-driven job displacement, ethical risks, bias amplification, and the erosion of human agency in critical decision systems. As a tool, it enables unprecedented operational efficiency, analytical capability, and productivity through data-driven insights and automation of complex cognitive tasks. Most critically, as a transformation partner, AI represents a co-evolutionary force reshaping business models, empowering human creativity, and redefining the nature of work and leadership. This triadic lens underscores the importance of governance, transparency, and strategic collaboration to ensure that AI's transformative potential aligns with human values and institutional resilience. The future of AI integration will be determined not by the technology alone, but by how thoughtfully societies balance these dimensions in pursuit of trusted, human-centric innovation.

When RAG systems fail, the first instinct is often to blame the model. In practice, weak answers are usually caused by poor evidence flow: the wrong chunks are retrieved, the question is interpreted too narrowly, or the final context is too thin to support a useful response.

This session breaks down seven practical ways to improve RAG answer quality in an AWS-native architecture before reaching for fine-tuning or a full rebuild. Using a real-world transcript-grounded assistant as the case study, I will cover chunking strategy, metadata design, query rewriting, multi-query retrieval, reranking, synthetic summary chunks, and citation mapping.

The focus is not on flashy demos. It is on the architectural decisions that make RAG systems more accurate, more explainable, and easier to debug in production. Attendees will leave with a practical framework for diagnosing weak RAG behavior and improving quality using AWS-native services and patterns.

Everyone's saying MCPs are dead and you should just use bash. They're not wrong, about 80% of the time. But what about the other 20%?

This talk makes the honest case for when MCPs are still the right answer: SaaS integrations with no CLI, enterprise security boundaries, dynamic tool discovery, and structured communication that doesn't require parsing stdout with regex. We'll cover practical lessons for building MCP servers that actually work well with LLMs: keeping tool counts low, designing around outcomes instead of endpoints, and writing error messages that let agents self-correct.

Then we flip it around. If you're on team "just use bash," your CLI better be designed for agents, not just humans. We'll walk through what makes a CLI agent-ready: structured JSON output, meaningful exit codes, JMESPath filtering, dry-run modes, and shipping the prompt alongside the binary.

Building a multi-agent AI system is the easy part. Running it in production is where the real lessons begin. In this session, we share hard-won insights from building a six-agent AI assistant for a fashion retailer, on AWS Bedrock AgentCore. We'll walk through the architecture decisions that worked and the ones that surprised us. Whether you're designing your first multi-agent system or scaling one to production, this session gives you a practical framework for making it work — and a clear-eyed view of what can go wrong.

The AWS Community has a place for every builder. Whether you like creating content, organizing tech events, or just showing up to learn. Pick your level of commitment, choose what excites you (events, content, code), and jump in wherever feels right. Your community, your adventure.