Rework demo READMEs and web pages: why each provider, one shared look

[SamyMe]

Summary

This builds on #8 (news-to-video). Merge #8 first; this PR's diff will then show only the rework commit.

READMEs

• Structure: all three demo READMEs now open the same way: pitch, flow, At a glance (providers, time and cost per run, how it runs, endpoints), then Why these providers.
• Why these providers: a table giving each provider's role in that demo and why it's a great fit. The reasons are backed by numbers we measured with real calls, and vendor claims are attributed.
  • Tavily: a search API built for AI agents that returns cleaned, ranked page content. About 1 s and $0.008 per search. In Upstream Watch it surfaced real CVEs.
  • Nebius: open-weight models (Qwen3-235B, gpt-oss-120b) on Token Factory, with reliable tool calling and strict JSON output, for under $0.001 per call.
  • Gradium: the voice specialist from the Kyutai research lab team, EU-hosted on Eden AI. Accurate transcription of browser recordings and a natural voice. Its reported Coval latency results are attributed and clearly separated from the non-streaming path this demo uses.
  • Pruna: fast video. A 5 s vertical clip with sound for $0.10, rendered in parallel. Its Design Arena ranking is attributed to Pruna.
• Root README: a new Providers in this cookbook table.
• Voice article: a "Why Gradium, Tavily and Nebius" section.
• Stale details fixed: timings, the output folder naming, the article's vendor count, and how the voice demo now serves audio.

Web pages

• One shared look for both demos:
  • an Eden AI Cookbook header;
  • provider chips showing each provider's role;
  • a live pipeline strip whose stages light up with their provider and time;
  • a footer on the one-key, one-bill setup.
• Voice: the pipeline labels follow the dropdowns, so choosing Linkup renames the Search stage. A failing stage is marked, and malformed source URLs can't break rendering.
• News: the step checklist becomes the same pipeline strip, and the failing stage is marked.

Upstream Watch has no web page: it runs as a GitHub Action, so its README is the only change there.

Testing

Tested in Chrome against the mock and fake Eden AI backends, so no billed calls:

• Voice page: typed flow through all four stages, live label swaps, and the text-to-speech failure path, where the answer stays on screen and Speak is marked failed.
• News page: all five stages, the three clip slots, the final video and the cost table.
• Playback verified in a foreground tab, which earlier runs couldn't check:
  • the news clips and the final video play;
  • the voice answer plays: a real Gradium streaming WAV, repaired by the backend, with duration 10.56 s and the playhead advancing.

🤖 Generated with Claude Code