- Why multimodal is a shift: video, images, and audio are becoming first-class AI data, but most stacks still treat them as opaque files, leading to brittle pipelines and high maintenance overhead - Pixeltable’s core idea: a more “database-like,” declarative table/view model with native media types, built-in transforms + model inference, and strong lineage/observability - What we demoed + discussed: ingesting video by URL, extracting/transcribing/chunking, building embeddings for semantic search, and an auto-crop example, plus Q&A on flexibility (UDFs/models), private deployment, and scaling direction

In this session, we explored why agentic AI can be risky: models can sound confident while hallucinating, skipping required steps, or overstepping their authority. We covered semantic authorization, granting permissions based on meaning, context, and output quality, and demoed a control-plane pattern that enforces pre-checks, calibrates thresholds against known-good responses, and records a full audit trail tied to a human owner for real workflows.

During the discussion with Span, we talked about a challenge many eng leaders are navigating: AI is increasing coding throughput, but overall delivery impact can feel flat as code review, rework, and quality bottlenecks emerge. We also discussed what it would take to measure AI impact day-to-day: going beyond vanity metrics and surveys to instrument the full SDLC, surface actionable insights, and tie adoption to quality and cost.