Agentic AI Conference

The Last Mile of OCR/LLM-Based Document AI

Hands-On with Agentic Document Extraction

OCR performs well in benchmarks, but real-world document AI lives in the long tail: large tables, old scans, mixed-language documents, handwriting, and complex layouts. Even top models struggle here. This workshop demonstrates how LandingAI's Agentic Document Extraction (ADE) goes beyond OCR and parsing to handle real-world enterprise workloads effectively.

• Explore the core pillars of Agentic Document Extraction and understand how they address real-world document challenges.
• Build end-to-end document processing pipelines using ADE API and SDK for structured, reliable outputs.
• Use Skills to automate tasks and enable coding agents to generate solutions for complex document workflows.
• Apply ADE to enhance LLMs' ability to process large tables, scanned documents, and mixed-language or handwritten content.

Ankit Khare

Developer Relations Specialist

Reinforcement Learning with Human Feedback and GRPO

How Do You Get Your LLM To Do Math?

This session tackles the challenge of making Large Language Models (LLMs) not just fluent in text, but capable of reasoning in math and code. Talking is imitation – learning by example – but logical tasks require learning by objective. That's where RLHF comes in.

• Understand why LLMs' capabilities in reasoning, math, and coding emerged later than their language abilities.
• Learn what Reinforcement Learning with Human Feedback (RLHF) is and why it's essential for fine-tuning models.
• Dive into Group Relative Policy Optimization (GRPO), a popular method for optimizing models on logical tasks.
• Explore practical applications of RLHF and GRPO to make LLMs more reliable in reasoning-intensive scenarios.

This serves as a theoretical companion to Chris McCormick's workshop. Attendance to Chris's workshop is not required, but is recommended for anyone who is interested in a practical coding example of GRPO.

Luis Serrano

Founder & Educator

From Isolation to Trust

Building Secure Sandboxes for Autonomous AI Agents

This session tackles the real problem with autonomous AI agents: speed is great, but unconstrained access is a liability. You want the performance. Not the security nightmares. Enter Docker Sandboxes — a new primitive designed to let agents operate in a restricted environment with controlled filesystem access, network permissions, and secret injections.

• Understand the typical ways AI agents go wrong when given unrestricted system access — and why basic isolation isn't enough.
• Learn how Docker Sandboxes work as a controlled environment for autonomous agent execution.
• Walk through filesystem restrictions, network controls, and secure secret injection in practice.
• Build a workflow for running AI agents you can actually trust in production.

Oleg Šelajev

Developer Relations

From Retrieval to Orchestration

Building Intent-Driven Agentic Context Engineering Systems

Most agentic RAG systems break in production because they treat every query the same. Real users ask multi-part questions, shift intent mid-conversation, and expect systems to keep up. This workshop shows you how to move beyond single-shot retrieval and build systems that actually handle real-world complexity.

• Understand how vector databases store and retrieve context, and why hybrid search and metadata filtering outperform dense-only approaches in production.
• Learn to decompose complex, multi-intent queries into targeted sub-queries for more accurate and reliable retrieval.
• Build an intent-based orchestration layer where specialized agents handle distinct tasks with dedicated tools and strategies.

Chris McCormick

AI Researcher and Educator

Advances in Agentic AI for Healthcare

Hands-on with Multi-LLM Clinical Applications

This session goes beyond conventional AI demos. While LLMs excel in controlled scenarios, real-world healthcare data is messy: electronic health records (EHRs) are fragmented, multi-modal, and require contextual reasoning across patient histories. This talk demonstrates how agentic AI systems, built on LangGraph agents orchestrating multiple LLMs within SambaNova's SambaStack runtime, tackle these real-world challenges.

• See how intelligent agents integrate specialized tooling to generate and execute Cypher queries on a Neo4j graph database for secure, high-fidelity EHR analysis.
• Explore multi-turn conversational workflows that support complex clinical queries and similarity searches across patient profiles using node embeddings.
• Understand how dynamic runtime model selection across co-deployed LLMs balances latency, throughput, and task-specific performance.
• Learn how on-premise SambaStack deployments can reduce total cost of ownership while maintaining data privacy, security, and production-grade reliability.

Varun Krishna

Senior Principal Solutions Engineer - Generative AI

From GRPO Theory to Practice

Hands-On GRPO Training Pipeline To Fine-Tune A Qwen Model

A practical companion to Luis Serrano's April 8 workshop on RLHF and GRPO, this session moves from theory and equations to running a stable training loop. We'll walk through a GRPO training pipeline to fine-tune a Qwen model on arithmetic tasks, focusing on the design decisions that determine whether reinforcement learning trains successfully.

• Walk through a GRPO training pipeline covering generation, reward calculation, and optimization.
• Understand the key hyperparameters behind stable RL training, including prompt batch size, group size, and learning rate.
• Learn how to interpret RL metrics such as the standard deviation of rewards.
• Explore techniques for diagnosing instability and reward hacking using Weights & Biases.
• See how GRPO systems scale using tools like vLLM and multi-GPU training.

Chris McCormick

AI Researcher and Educator