Tech Blog
Engineering posts and talks our team has published on external venues

RAG ๋ค์์ Knowledge Layer: ๊ฒ์ ๋ค์ ๋จ๋ 4๊ฐ์ง
Finding the right document is not the same as knowing the right answer. What Microsoft Foundry IQ, AWS's semantic layer and Snowflake Horizon Context solve under different names, broken into four problems: consistency, freshness, permissions and provenance โ plus a field fix using Korean bigram indexing.

AI ์์ด์ ํธ๋ ์ ์๋ก ํต์ ํด์ผ ํ ๊น? A2A์ ๋ชจ๋ ๊ฒ
Agent competition is shifting from "how smart is one agent" to "how do different agents meet, divide work, exchange state, and verify results." A breakdown of the A2A Protocol v1.0 โ Agent Card, Task Lifecycle, capability discovery, modality negotiation โ its relationship to MCP and ACP, and the Linux Foundation backing of 150+ organizations.

ํด๋ก๋ ์ฝ๋ ์์ค ์ ์ถ์์ ๋ฐฐ์ฐ๋ ์์ด์ ํธ ๊ตฌ์กฐ
Agents differ in structure, not the model. From the leaked Claude Code internals: 42 tools across 5 permission tiers and an MCP/Statsig/Sentry observability layer.

ํ๋ค์ค ๊ฒฝ์์ ์์, ์ opencode์ OMO์ผ๊น?
"The harness, not the model, decides." How opencode + OMO, proven across 30 days and 2.47B cached tokens, lifted long coding tasks to completion.