Latest blog posts by Paweł Szczepanik

An internal AI platform is the small set of components every AI team would otherwise rebuild: gateway, prompt registry, evaluation harness, templates. Built once, owned as a product.

RAG evaluation splits into two questions: did retrieval find the passage, and did the answer stay inside it. Metrics, golden sets, LLM judges and the gate that catches regressions.

Gemini 3.8 Live Extended Thinking is not AGI, but it is a real step toward a voice assistant that keeps working while you talk. Benchmarks, pricing, limits and a 3-minute voice sample.

Live webinar on September 30, 3:00 PM CEST: how AI-powered pricing optimization helps retailers protect margin and clear inventory across 50,000+ SKUs.

DS Stream joins NASSCOM Agentic AI Confluence 2026 in Bengaluru on September 24. Meet us to talk decision intelligence for retail and CPG.

AI governance framework built into the delivery pipeline: blocking, advisory and human gates, runtime guardrails, and release evidence generated by every build instead of an audit project.

AI factory reference architecture as logical layers, not GPU racks: five layers, two planes, the contract each one exposes, and one design mapped onto any major cloud.

Model monitoring in production has to catch decay, not just drift. Four signals, the label latency problem, thresholds that hold, and alerts people act on.

AI factory cost is driven by integration, data work, governance and run rate, not GPUs. The seven budget categories, build versus run rate, and cost per use case.

An AI center of excellence sets standards; an AI factory delivers them. Here is where each model breaks and what has to exist before the CoE can step back.