AI Engineer World's Fair 2026 Wiki

AI Engineer World's Fair 2026 ran June 28-July 2, 2026 in San Francisco, with main programming at Moscone West. This standalone wiki turns the official schedule, public AI Engineer YouTube evidence, livestream transcripts, and extracted slide data into an interlinked conference intelligence map.

The wiki is schedule-first. Official session and speaker pages are treated as the backbone; YouTube videos, transcripts, OCR, and reconstructed slides are supporting evidence unless a video is confirmed as an exact session recording.

Quick Access

• index — full wiki index and section guide.
• talk video transcript map — official talks mapped to related YouTube videos and transcript status.
• slide library — extracted full-frame slide decks from related videos.
• reconstructed slide library — cleaner slide-region crops produced by the advanced slide scanner.
• dense slide library — high-density slide views for fast inspection.
• worldsfair 2026 livestreams — livestream uploads, transcript status, and slide extraction.
• advanced slide logic page audit — RapidOCR and reconstructed-slide audit.
• source boundary — corpus rules and evidence confidence.

Corpus

• 560 official schedule sessions generated as talk pages.
• 553 people pages generated from the official speaker roster plus schedule-only names.
• 340 company pages generated from speaker affiliations.
• 117 resource pages for official data, source boundaries, YouTube evidence, and processing audits.
• 310 slide pages, including full-stage frames, reconstructed crops, and dense slide views.
• 11 topic pages seeded from repeated World’s Fair themes.
• 5 event/day pages for the conference flow.

Evidence Layers

• Official schedule and speaker data establish dates, titles, speakers, organizations, tracks, and the event structure.
• Public AI Engineer YouTube metadata identifies related videos by speaker and title matching.
• YouTube transcripts are used where available; local Whisper fallback is part of the ingestion path when captions are missing.
• Slide evidence comes from video frame extraction, OCR, RapidOCR repair passes, and OpenCV crop reconstruction.
• Topic pages synthesize repeated language across titles, transcripts, slide text, and session clusters.

Slide Intelligence

Slides are first-class material in this wiki because many World’s Fair videos are stage recordings rather than clean deck exports.

• 107 extracted video slide decks are tracked in slide library.
• 1,842 slide/frame images are embedded in deck pages.
• 107 reconstructed companion decks are tracked in reconstructed slide library.
• 1,729 reconstructed slide-region images are saved under wiki/assets/reconstructed-slides/.
• 1,826 extracted frames have non-empty local OCR text.
• 291 weak OCR frames were improved with local RapidOCR/ONNX after the first Tesseract pass.
• 115 weak frames were classified as unreadable or non-slide cutaways during the RapidOCR audit.

Current Theme Map

• coding agents — coding agents as tools, teammates, infrastructure, and organizational pressure.
• software factories — self-driving software, agentic SDLC, code review, CI/CD, and production loops.
• agent evaluations — evals, harnesses, observability, quality gates, and agent performance review.
• agent memory — memory, context engineering, knowledge agents, and long-horizon work.
• agent security — permissions, provenance, identity, governance, and agent supply chain risk.
• agentic search — search, retrieval, context graphs, GraphRAG, and agent web navigation.
• voice agents — realtime voice systems, speech-to-speech models, and voice UX failures.
• inference engineering — inference engines, model serving, open weights, optimization, and deployment tradeoffs.
• mcp — MCP servers, tool layers, agent interfaces, and operational governance.
• ai sandboxes — sandboxes, browsers, environments, and safe agent execution.
• autoresearch — research agents, data quality, parameter golf, and scientific workflows.

Reader Contract

This is a clean AIE-specific conference wiki, not a mixed local workspace. It intentionally avoids cross-project notes, private recordings, diary material, and unrelated local vault content. Pages should make source confidence clear: official schedule facts are canonical, related videos are supporting context, and slide/OCR text is best-effort until manually reviewed.