Event Resources
Slides, code, and resources from my talks
14 Tool Calls for One Booking: How to Stop Agent Loops
AI agents get stuck calling the same tool repeatedly, burning tokens and time. DebounceHook detects duplicates, clear states stop loops, and LimitToolCounts enforces hard ceilings. Drop call counts by 85%.
Agent Speedrun: Idea to Code to Deploy to Production
Hands-on workshop: build one customer service agent end to end with the open-source Strands Agents SDK, then take it to production on Amazon Bedrock AgentCore. Seven progressive modules, from the agent loop to deploy.
Break Your Agent with Chaos Testing, Then Fix the Harness
A demo is one clean session. Production breaks the same agent in four ways a better prompt cannot fix. Diagnose them with chaos testing, then fix each one in the harness around the model: validate before the write, gate the action at the tool boundary, verify against ground truth, and let the agent write its own tools.
Catching Hallucinations with Multi-Agent Validation
Single AI agents hallucinate non-existent data. Multi-agent validation uses Executor, Validator, and Critic roles to cross-check outputs and catch errors before users see them.
Context Engineering: Stop Agents from Choking on Their Own Data
AI agents fail silently when tool outputs exceed context windows. The Memory Pointer Pattern stores large data in agent state and returns 52-byte pointers instead of 214KB payloads.
How to Stop AI Agent Hallucinations: 5 Targeted Fixes
AI agents hallucinate in five distinct ways. Each failure needs a different fix. Five targeted defenses validated across 8 test scenarios with zero hallucinations.
MCP Tools Keep Timing Out? The Async Pattern That Fixes 424 Errors
MCP tools calling external APIs hang for minutes, killing AI agents with 424 errors. The async handleId pattern fixes this: start_long_job returns a job ID immediately, check_job_status polls results.
Multimodal AI Agents with Long-Term Memory
Multimodal agents process video, images, and text, but forget everything when the conversation ends. Build agents with scalable chat memory using MCP servers and a managed vector store.
Multimodal RAG: Video Search Without the Pipeline
Video RAG used to require frame extraction, audio separation, multiple embeddings, and complex orchestration. Multimodal models changed everything.
Orchestrating the Pipeline with Strands Agents
Build a multi-agent video processing pipeline in one hackathon session using the Strands Agents SDK — transcription, summarization, and indexing agents working together.
Real-Time Voice Agent
Real-time voice agents that access custom knowledge bases to create adaptive learning experiences with dynamic feedback and persistent progress tracking.
Reduce AI Agent Costs with Semantic Tool Selection
AI agents waste tokens sending all tool descriptions on every call and pick wrong tools as they scale. Semantic tool selection with FAISS filters to the top 3 relevant tools per query, cutting tokens by 89% and improving accuracy.
Ship It: From Agent Demo to Production in Minutes
Your agent demo wowed the team. Six months later, you are still rewriting it for production. Learn how to add memory, monitoring, and auto-scaling to any agent.
Stop AI Agent Hallucinations: 5 Techniques + Production Patterns
Hands-on workshop covering Graph-RAG, semantic tool selection, multi-agent validation, neurosymbolic guardrails, and agent steering with production deployment on Amazon Bedrock AgentCore.
Stop the Loop That Hallucinates and Burns Tokens
Your agent fabricates data and burns tokens on the same ungoverned loop. A better prompt fixes neither. Ground answers in a structured source, route tools by meaning to cut wrong picks and cost, enforce rules it cannot bypass, and block repeat calls.
The Infinite Context Window Is a Myth: Context Engineering for AI Agents
Agent systems break when tools return large volumes of data and the context window overflows. The Memory Pointer Pattern externalizes large outputs and guides workflows with lightweight references, in single-agent and multi-agent setups, with real metrics.
The Managed Harness: Define an Agent, Skip the Orchestration Code
Every agent needs an orchestration loop, and building it for production means concurrency, isolation, identity, memory, and observability. The managed harness turns that work into configuration: declare model, tools, and instructions, and the platform runs the agent in an isolated microVM.
When Prompts Fail: Enforcing Business Rules in AI Agents
You wrote 'max 10 guests' in the prompt. The agent booked 15. LLMs treat rules as suggestions. Learn how to enforce business rules at the code level using Python hooks.
When RAG Hallucinates Numbers: Graph-RAG for Precise Answers
Traditional RAG agents hallucinate when they need to count, aggregate, or detect out-of-domain queries. Graph-RAG with knowledge graphs eliminates these statistical hallucinations through precise queries.
Why AI Agents Forget Everything
Your AI agent forgets who you are after every conversation. Three memory patterns fix this: persistent state, core memory, and semantic retrieval.
Your Agent Works on Localhost. Now Ship It
Anti-hallucination techniques work in notebooks but break in production. Deploy GraphRAG, semantic routing, guardrails, and multi-agent validation with database-backed steering rules.
Your Agent's Context Window Is Full. Now What?
Hands on workshop: build three fixes for context window failures using four context engineering strategies (externalize, select, compress, isolate). Memory pointer in single and multi agent setups, debounce for loops, async handle for slow tools. Real metrics.
Your AI Agent Isn't Crashing. It's Bleeding Tokens
Your AI agent does not crash; it silently wastes tokens. Context overflow, MCP 424 errors, and reasoning loops. Three fixes with real metrics: 7x token reduction, 17s to 1.7s, 14 calls to 2.
No talks match that filter. Try another tag or search term.