Agentic AI, LLM engineering, and the models behind modern automation — multi-agent systems, LLM evaluation and comparisons, RAG, fine-tuning, AI infrastructure, security, and production AI engineering.
From the blog
Mcp http vs stdio: mCP 1.0 supports stdio and Streamable HTTP transports for different deployment shapes. When each one is the right pick and how authentication differs between them.
Qdrant and Milvus took different architectural bets. Head-to-head on throughput, latency, ops complexity, and where each one wins in serious production deployments today.
A practical guide to AgentKit 1.0's evals harness — golden traces, LLM-as-judge, regression gates, and how to ship agent updates safely in 2026.
By April 2026 agent runtime sandboxing has three credible options — gVisor, Firecracker microVMs, and WASM — with different latency, capability, and tool-call tradeoffs.
Four April 2026 incident reports show the same data-exfil pattern via agents — a poisoned document instructs the agent to read a sensitive store and post to an attacker URL.
Anthropic and the MCP working group published a draft registry + signing spec in April 2026, with sigstore-style attestation and a curated trusted publisher list.
SentinelOne's Purple AI April 2026 release adds multi-tenant hunting and an MCP server for SOC tools, letting analysts query telemetry from any MCP-aware client.
Every LLM call should emit a span with model, input tokens, output tokens, and cost — collected via OTel and aggregated in ClickHouse or Langfuse. We show the schema, the per-tenant cost cap pattern, and the daily founder digest.
NVIDIA folded Triton into Dynamo in 2026. Self-host Whisper, NeMo, Riva, and NIM speech microservices on L40S/B200 with gRPC streaming. Production blueprint for HIPAA-locked voice.
