Open Source Summit Korea 2026

The team shipped semantic search. It worked. Then came 100M vectors, a RAM bill that tripled overnight, filtered queries silently returning zero results, and a proposal to "just add Pinecone alongside Postgres."

This talk is for engineers who've already built the thing and now need to scale it without burning the budget or the team. We'll do the RAM math that should happen before any architecture decision, walk through quantization strategies that deliver 32x compression with 95%+ recall, show why filtered search is the #1 silent production failure, and lay out when DiskANN, hybrid BM25+vector search, or a specialized vector DB actually makes sense — and when it doesn't.

Live demos included. Decision matrices to take back to the team on Monday. No hype, just trade-offs in plain English.

Open-source agents like Hermes and OpenClaw can reason and use tools, but edge deployment is still hard - not the model, the compute underneath: limited memory, tight power budgets, no ops team. Democratizing agentic AI means fixing the compute layer. This deep dive covers GPU and memory slicing at the edge: carving one device’s unified memory so multiple agents run concurrently on hardware as small as an NVIDIA Jetson, and time-sharing when demand exceeds memory.

We compare Jetson-class GPUs with NPUs from Axelera and DeepX - higher performance per watt - and what it takes to make them schedulable. HAMi, the CNCF Incubation project for hardware-agnostic GPU virtualization, is that layer: one scheduling plane across heterogeneous accelerators, fine-grained memory slicing, open source. A blueprint for edge AI without a cloud budget.

Trusted OSS is an open-source self-certification kit that guides any organization from zero to ISO/IEC 5230 (license compliance) and ISO/IEC 18974 (security assurance) conformance using AI agents.
No prior expertise required.

Built by the OpenChain Korea Work Group and released under CC BY 4.0, the kit features:

• AI agents (Claude Code) that auto-generate company-specific compliance artifacts: OSS policy, SBOM, vulnerability response procedures, training curriculum, and conformance declaration
• DevSecOps pipelines (SAST, SCA, secret detection, IaC) ready to drop into any CI/CD environment
• A 5-level AI Coding Governance Maturity Model — from ad-hoc prompting (Level 1) to AI-augmented defense (findings-driven review, AI fuzzing) and continuous auto-remediation (Level 5).

In this session, I'll walk through how any team can go from no compliance process to a fully documented, self-certifiable program in hours, not months. I'll also share how we're using AI to close the compliance skills gap across Korean enterprises and SMEs — making OpenChain certification accessible to all.

Attendees leave with a working toolkit they can clone and run today.

AI coding tools can generate patches quickly, but maintainers still cannot merge a PR on generated code alone. They need evidence, what was reproduced, what changed, how it was validated, and whether the contributor actually reduced review work instead of shifting it downstream.

In this session, I draw on recurring patterns from reviewing and maintaining Argo CD contributions to show why many AI-assisted PRs stall and what makes others move quickly. I present a simple four-stage framework for review-ready OSS contributions, Set up the environment, validate the problem before changing code, harden the fix with tests and checks, and submit a PR with enough context and proof for efficient review. Attendees will leave with practical guidance they can use immediately to improve contribution quality and reduce maintainer burden.

Building AI agents is easier than ever with open source tools, but ensuring their reliability in production remains a major challenge. Unlike traditional software, AI agents are non-deterministic, making simple pass/fail testing insufficient.

This talk introduces a practical approach to evaluation and observability for AI agents, combining open source tools such as TruLens with agent architectures inspired by AgentGPT.

We will demonstrate how to instrument agent workflows, capture execution traces, and implement evaluation metrics such as faithfulness, tool selection accuracy, and answer relevancy. Attendees will also see how to visualize agent behavior and identify failure points across retrieval, reasoning, and generation layers using a lightweight dashboard.

Finally, we show how to build a feedback loop to iteratively improve agent performance, and share a reference implementation (GitHub) that can be reused with different agent frameworks.

Building AI agents is easy, but ensuring their reliability is hard. Because agents operate in open-ended, non-deterministic environments, traditional Test-Driven Development (TDD) falls short. The solution is Eval-Driven Development (EDD), a paradigm that embeds continuous evaluation throughout the agent lifecycle.

This talk explores how to operationalize EDD. First, we examine the critical role of observability. Evaluating just a final output is insufficient for multi-step workflows. We show how OpenTelemetry (OTel) compatible tracing exposes internal reasoning and tool usage for granular debugging.

Next, we tackle scaling evaluation via LLM-as-a-Judge. Since uncalibrated judges often miss domain-specific nuances, we demonstrate aligning custom judges with human experts. Using open-source MLflow, attendees will learn to capture traces, collect expert feedback, and use alignment optimizers to create expert-aligned evaluators.