AI Engineering Talks — Classified & Distilled

English · 繁體中文 · Live site ↗

A systematic read-through of 159 conference and YouTube talk notes on AI engineering, sorted into 9 thematic categories and distilled into short, skimmable key-point summaries — plus 9 cross-cutting insights drawn from the entire corpus.

The published page is a single, self-contained index.html — no server, no external dependencies, just open it in a browser. That file is now generated from small, composable sources under src/ by a tiny build script (build.mjs), so you edit focused partials instead of one 44k-line monolith. CSS and JS are inlined at build time, keeping the output dependency-free.

The site ships in English and Traditional Chinese. A language toggle (EN | 中文) sits just above the View source on GitHub button; the first visit follows your browser's language (any zh* preference → Chinese, otherwise English) and your choice is remembered. The build emits two self-contained pages — index.html (English) and index.zh.html (Traditional Chinese) — and the toggle switches between them while preserving your place on the page.

Highlights

• 159 talks distilled into short key-point summaries.
• 9 thematic categories (A–I) with a category-distribution overview.
• 9 cross-cutting insights synthesized across all talks.
• Dual evidence for every summary:
  • Click a talk title (or the 📄 Full notes button) to expand the complete original notes, embedded directly in the page.
  • Follow ▶ Source video back to the original YouTube talk.
• Personal notes — inside a talk's full notes, select a sentence (or, on a phone, tap a sentence — tap again to extend across more sentences) and choose ★ Save as note to highlight it. Saved notes are collected in a Your Notes section grouped by talk and persist in your browser.
• Fully self-contained output — all notes are rendered and embedded, so the published index.html depends on no external files or network connection.
• Composable sources — the page is built from small partials (src/), so it's easy to edit and extend without touching a giant single file.

Categories

#	Category	Talks
A	BI / Analytics / Semantic Layer	15
B	Agent Evaluation & Observability	16
C	Agent Architecture, Reliability & Productionization	17
D	Agent Security & Identity	8
E	Context / Memory / RAG	18
F	Data Infrastructure	18
G	Model Training & Inference	21
H	AI Coding & AI-Native Engineering	12
I	Product Strategy & Business	6

Each talk is assigned a single primary theme.

Cross-cutting insights

1. The semantic layer is being redefined — sinking from BI tools down into "context that agents consume."
2. The key to reliable text-to-SQL is grounding and data modeling, not a bigger model.
3. Evals move from "gut feel" to data-driven engineering.
4. From PoC to production: reliability is systems engineering, not a model problem.
5. Context engineering and memory decide whether an agent uses the right data.
6. Data infrastructure is being reshaped for AI / agents.
7. Security and identity are prerequisites for agents to access enterprise data.
8. The future of BI and the product data flywheel.
9. Small, specialized models / agents beat big and general-purpose ones.

Usage

The two pages are build artifacts and are no longer committed, so build them from source first, then open one in a browser:

npm run build          # emits index.html + index.zh.html

# macOS
open index.html
# Linux
xdg-open index.html
# or serve locally, then visit http://localhost:8000
python3 -m http.server

The always-on version is the live site — see Deployment.

Content pipeline

New talks reach the site through a lightweight discover → pick → draft → review → release pipeline. The site stays backend-free — the "database" is git-tracked JSON plus GitHub Actions — and a GitHub Issue is the control panel you drive it from, so you can even pick from your phone:

channels.json ─poll (cron)─▶ queue.json (develop) + 📥 review Issue
                     Issue: triage.mjs ranks each pending video ⭐/🤔/⏭️ (OpenAI) + ✅/❌ boxes
   ② tick ✅/❌ + 🚀 submit ─queue-control─▶ approved   (one commit on develop)
   ③ gen-note --pr (transcript → Claude draft + 中文 + category) ─▶ PR (linked on the Issue)
   ④ review PR ─merge─▶ develop ─⑤ Release (one click)─▶ main ─▶ Vercel build + deploy

• tools/poll.mjs — pull each channel's RSS into queue.json on develop (scheduled by .github/workflows/poll.yml)
• tools/triage.mjs — score the pending queue with the OpenAI API (gpt-4o-mini) and render the review Issue as a ⭐/🤔/⏭️ control panel with checkboxes
• Pick on the Issue — tick ✅/❌ then the bottom 🚀 submit; queue-control.yml applies the batch to develop as one commit (via tools/queue-apply.mjs). Offline alternative: tools/review.mjs, a local approve/reject UI.
• tools/gen-note.mjs --pr — transcript → house-style note + Traditional-Chinese translation + category suggestion; opens a review PR and records the batch → PR back on the Issue
• Branch model: queue.json lives on develop; main is a pure release branch, published with one click via the Release (develop → main) workflow.
• Two human gates: which videos to keep, and reviewing each generated note in its PR

Day-to-day runbook — the steps you actually perform — in OPERATIONS.md (繁體中文). Design notes and phase-by-phase status live under docs/ (PRDs + todos).

Deployment

Deployed on Vercel with Git integration — every push/merge to main runs npm run build and redeploys; pull requests get their own Preview Deployment. Publish reviewed content from develop with one click via the Release (develop → main) workflow (Actions → Run workflow) — it opens the release PR, waits for CI, and merges.

• Live: https://ai-talk-notes.vercel.app
• The pages (index.html / index.zh.html) are not committed — Vercel builds them from source (vercel.json sets buildCommand: npm run build).
• Pure-data commits — the poll bot / the Issue control panel updating queue.json on develop — skip the rebuild via Vercel's Ignored Build Step, so only real content changes trigger a deploy.

Project structure

index.html        # generated English page (build artifact — git-ignored)
index.zh.html     # generated Traditional Chinese page (build artifact — git-ignored)
build.mjs         # renders src/* into both pages (inlines CSS + JS)
package.json      # `npm run build`
tools/            # content pipeline (poll · triage · review · gen-note · queue-apply) + content-check + i18n-check
src/
  head.html       # document head (minus styles)
  styles.css      # all page styles
  partials/       # hero, nav, footer, lang-toggle
  sections/       # overview.html, themes.html; cat-*.md (English card text + color/docs)
  notes/          # shell.html + doc-*.md (English notes) + order.json
  scripts/        # modal, reading-progress, notes, nav-scrollspy, lang
  i18n/zh/        # Traditional Chinese content (notes/*.md, sections/cat-*.md, HTML mirrors)

All talk notes and category cards are authored in Markdown and rendered into language-agnostic HTML; see src/README.md for the full layout and CONTRIBUTING.md for how to add or translate one.

Development

index.html is a build artifact — edit the files under src/ instead, then regenerate:

npm run build   # or: node build.mjs

The build only concatenates, inlines, and renders (no dependencies to install), and the result is reproducible. Requires Node.js. One run emits both index.html and index.zh.html.

All content is Markdown. Talk notes (notes/doc-*.md) and category cards (sections/cat-*.md) are authored as Markdown / flat text and rendered into language-agnostic HTML — English under src/, each translation under src/i18n/<locale>/. Untranslated pieces fall back to English. After building, run node tools/i18n-check.mjs to verify the two pages stay structurally identical.

Contributing

Contributions — new talks, corrections, and translations — are welcome. All content is authored in Markdown, and one build renders both language pages from it. See CONTRIBUTING.md (繁體中文) for the content model, the Markdown formats, how to add a talk, and how to add or improve a translation.

Method & evidence

• Data source: all 159 Markdown talk notes, fully rendered and embedded in the built page (no dependency on external .md files).
• Dual grounding: every summary links to both the full original notes (expandable in-page) and the source YouTube video.
• Classification: a single primary theme per talk across the 9 categories (A–I); every insight is drawn from the full corpus.

Credits

This project is based on cyyeh/ai-talk-notes (MIT License), with additional features by yeha98555.

License

Released under the MIT License. © 2026 cyyeh, yeha98555.