{
  "video": {
    "id": "jebp4V0vh30",
    "title": "Agentic Sites: Building Hyper Personalized Websites — Carlos Sanchez, Adobe",
    "duration": 1243,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "jebp4V0vh30",
    "title": "Agentic Sites: Building Hyper Personalized Websites — Carlos Sanchez, Adobe",
    "one_liner": "Adobe's Carlos Sanchez demos \"agentic sites\" — AEM Edge Delivery pages whose individual blocks are regenerated per visitor from a RAG index of the site itself, in ~1 second using Gemma 4 on Cerebras.",
    "summary": "Sanchez, a principal scientist on Adobe Experience Manager, argues that real-time hyper-personalized websites are now practical because inference got fast enough: the page has to assemble in 1–2 seconds or it costs conversions. Rather than generating whole pages (marketing brand guidelines forbid it), the system personalizes selected blocks — hero card, products, blog feed, navigation, CTAs — grounded in a RAG index built from the site's own content, driven by browsing signals bucketed into personas/intent types that marketers define in natural language. He shows continuous Promptfoo evals across models and providers scoring both accuracy and latency, a live coffee-equipment site generating a camping-focused page on the fly, an internal tool that turns any URL into an agentic site in under an hour, and a Google TV voice query producing a personalized page.",
    "key_points": [
      "Only blocks are personalized, not the whole page: \"if you talk to marketing people they have very strict brand guidelines\" — the entire site is used as a corpus and a RAG is built from it so generated content is grounded on existing site content.",
      "Evals run continuously with Promptfoo across many models and providers, scoring accuracy *and* speed; 15 prompts were curated for the example site, and the right model turns out to be highly site-dependent (size, vertical, commerce type), so the eval has to be re-run per site.",
      "Cerebras running Google's Gemma 4 (announced the week before the talk) averaged 1.1 seconds to generate a page; the next-best entry was 4.6 seconds, and the rest ran from 4 seconds up.",
      "In the live debug readout on the coffee site with Cerebras Gemma 4, LLM time was ~1 second at 2,200–2,300 tokens/second (the total-time figure he read aloud as \"164 seconds\" is inconsistent with the sub-second demo he had just run).",
      "You don't need a big model: the job is generating text and deciding which blocks to place and in what order, so a model that is merely \"good enough if it's fast enough\" wins for many of the sub-tasks.",
      "Browser-side signals — pages visited, time spent per page, queries — bucket the user into a category (the demo showed \"exploring\") and feed the LLM; the demo query \"coffee machine to prepare coffee while camping\" returned a page with camping-specific copy, coffee tips and two suitable machines.",
      "A \"For You\" recommendation page can be pre-generated and pre-fetched as the user browses, relaxing the latency requirement — but repeated regeneration as they navigate has real cost implications from multiple LLM calls.",
      "Architecture: browser signal layer → backend on Google and Cloudflare doing the LLM calls and RAG reasoning, plus a vector database and inference, with AEM Edge Delivery serving pages and static content at the edge.",
      "\"OfOneLabs\": an internal tool where you enter any URL and get an agentic site in under an hour — he did it for the AI Engineer conference site, where \"Europe AI conferences\" produced a focused page and another query produced a side-by-side conference comparison.",
      "Image generation on the fly is being considered (he cites the just-announced Nano Banana Light), but he's unsure marketers want generated imagery unless it's reliably on-brand.",
      "Final demo: a voice query to a personal assistant via Google returns a fully personalized page on Google TV — no phone or computer, just voice in the living room."
    ],
    "takeaways": [
      "Benchmark models on latency alongside accuracy, and re-run the benchmark per site/use case rather than picking one model globally — Promptfoo over a curated prompt set (15 in his example) makes this continuous.",
      "Budget 1–2 seconds for page generation and pick a provider/model accordingly; the 1.1s vs 4.6s gap between the top two entries decided the stack.",
      "Personalize blocks, not pages, and ground every generation in a RAG built from your own site content — that's what keeps output inside brand guidelines and out of hallucination territory.",
      "Move non-latency-critical generations (recommendation/\"For You\" pages) to pre-generation and pre-fetch as signals accumulate, and cost the repeated LLM calls that implies.",
      "Give marketers the controls: let them define the personalization strategy in natural language and decide how many intent/persona buckets exist, then close the loop with analytics."
    ],
    "topics": [
      "personalization",
      "rag",
      "evals",
      "inference-speed",
      "agents",
      "web",
      "cms",
      "marketing"
    ],
    "tools": [
      "Adobe",
      "Adobe Experience Manager",
      "AEM Edge Delivery Services",
      "Cerebras",
      "Gemma 4",
      "Google",
      "Amazon Bedrock",
      "Promptfoo",
      "Nano Banana Light",
      "Cloudflare",
      "Google TV",
      "OfOneLabs"
    ],
    "quotes": [
      {
        "text": "We don't want the whole site to be generated. I mean if you talk to marketing people they have a very strict brand guidelines.",
        "at": "02:40",
        "url": "https://www.youtube.com/watch?v=jebp4V0vh30&t=160s"
      },
      {
        "text": "we use the whole site as a corpus. We built a rack from the whole site. So what is generated is grounded on the existing site.",
        "at": "02:56",
        "url": "https://www.youtube.com/watch?v=jebp4V0vh30&t=176s"
      },
      {
        "text": "And you don't need a huge LLM to do this sort of work because you are generating text, you are deciding where to put blocks and how to organize the website, you don't need a lots of information for that.",
        "at": "08:01",
        "url": "https://www.youtube.com/watch?v=jebp4V0vh30&t=481s"
      },
      {
        "text": "This is something that we only dreamed about before.",
        "at": "16:04",
        "url": "https://www.youtube.com/watch?v=jebp4V0vh30&t=964s"
      }
    ],
    "words": 3354
  },
  "summary_url": "/#jebp4V0vh30",
  "transcript": {
    "html": "/transcripts/jebp4V0vh30.html",
    "txt": "/transcripts/jebp4V0vh30.txt",
    "vtt": "/transcripts/jebp4V0vh30.vtt"
  }
}