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
PCM16 at 24kHz burns 384 kbps. Opus at 32 kbps delivers indistinguishable quality. The bandwidth math says use Opus. The vendor API math sometimes says use PCM16. Here is when each wins.
What attestation level your AI calls actually receive, why Level A is the only one that lights up Caller Verified on the receiver's screen, and how to get there with a Twilio-style trust profile.
BigBlueButton, Zoom EDU, and Outschool all converge on the same WebRTC pattern for K-12 and higher-ed. Here is the architecture and the recording story for 2026.
Axum websocket: stand up a production voice bridge in Rust: axum for routing, tokio-tungstenite for the OpenAI Realtime socket, and broadcast channels for fan-out. Real working code.
Kafka 4.0 ditches ZooKeeper, ships share groups (KIP-932), and previews Eligible Leader Replicas. Here's how we fan out 50k AI call events per minute to analytics, billing, CRM, and ML training.
Stripe runs a remote MCP server at mcp.stripe.com and Sessions 2026 added the Link agent wallet. We unpack how an LLM goes from intent to settled payment with controls.
Use Modal's GPU memory snapshots, Rust-native container runtime, and warm pools to keep Moshi, Kyutai STT, and Llama under 800ms cold-start. Production blueprint for serverless voice.
Rerankers turn a noisy top-50 into a clean top-5. Cohere Rerank 3.5 hits SOTA on BEIR with 595ms latency; ColBERT v2 wins on cost. Here is how CallSphere routes between them per vertical.
NIST's AI Risk Management Framework is voluntary on paper and mandatory in practice. Federal contractors, banks, and SaaS buyers all map RFPs to it. Here is a 90-day Govern-Map-Measure-Manage rollout for voice AI teams.
