"I want to show you something before I explain anything. This is a live streaming platform — real film data, real AI behind every button. No slides, no mocks."

"Everything I'm going to demo in the next 30 minutes was added to this site without changing a single line of its backend code. The secret is one concept I'll show you in 60 seconds."

"Here's the one thing you need to understand. A Cloudflare Worker is a small piece of code deployed at the edge — in front of your existing site. When a request comes in, the Worker intercepts it first."

"Your origin server never knows. It gets a completely normal request and returns its normal response. The Worker reads it, modifies it, adds to it — and returns something better to the user. That's it."

"No Kubernetes. No redeployment. One command — wrangler deploy — and it's live in 330 cities in 10 seconds."

"Watch the bottom right corner of the screen."

"That button was not in the page before. The server returned exactly the same HTML it always does. A Worker intercepted the response and injected the chat widget using HTMLRewriter — a streaming HTML parser built into the Workers runtime."

"No backend deploy. No frontend PR. No waiting on engineering. 30 minutes of work."

"Instant AI recommendations, grounded in the catalog via AI Search with AutoRAG. The chatbot can answer questions about any film in the database — no manual FAQ writing, no stale content. It retrieves current data every time."

"No support ticket. No wait time. No human agent at 3am on a Sunday."

"Now let me show you something that hits your finance team immediately."

"Every image, every thumbnail your users load — AWS charges you to send that data out. They call it egress. It's typically 5–10% of cloud bills for any company with user-facing content."

"Cloudflare R2 stores the same content, serves it from the nearest Cloudflare PoP — and charges zero egress fees. Permanently. Not zero with a cap. Zero."

"But here's what makes it interesting for your engineering team — not just your finance team. When the Worker stores an image in R2, it writes custom metadata: exactly when that asset was first cached, its original file size, and where it came from."

"You can open the R2 dashboard right now and see every object with those three fields. You don't just know it's cached — you know when it was cached and from where. That's auditable infrastructure."

"Inception is in this catalog. Interstellar is in this catalog. The user typed exactly what they want to watch tonight — and got nothing. What do they do? They leave."

"Now the platform understands what they mean, not just what they typed. Inception. Interstellar. WALL·E. Same query, completely different results. The user stays, finds something, watches it."

"Your search engine wasn't rewritten. A Worker sits in front, embeds the query with AI, and returns semantically matched results. Your existing database is untouched."

"When someone in Thailand opens your subscription page and sees $9.99/month — they do the math. Is this available here? Why dollars? That friction kills conversions."

"Watch — this page shows local pricing automatically. Thai users see ฿350. Indonesian users see Rp162,000. Singaporean users see S$13. Same price point, right local context."

"No backend change. The Worker reads the visitor's country from request.cf.country — it's already there on every request, for free, from Cloudflare's network."

"If you produce video — lectures, films, training content — subtitles are a requirement in most markets. The standard approach is Amazon Transcribe at $0.024 per minute, plus an S3 bucket, IAM policies, and a batch pipeline."

"Workers AI Whisper does this at $0.0005 per minute — 48 times cheaper. One API call. No batch pipeline, no IAM policies, no S3 bucket."

"The browser splits the audio into 45-second chunks with 5-second overlap, sends 3 in parallel. Two AI passes clean the output — a Southeast Asian language model per chunk, then GPT-OSS 120B reads the full transcript and corrects domain vocabulary globally. You download a subtitle file ready for any video player."

"For 100 hours of new content per month — the difference between a $144 subtitle bill and a $3 subtitle bill."

"Speech-to-text, AI reasoning, text-to-speech — three models, one endpoint, under 2 seconds. No third-party SDK, no per-minute billing to a call centre platform."

"And if you want a live agent option — this opens a WebRTC video session. Visitor link for the customer, presenter link for your agent. No Zoom subscription. All routed through Cloudflare."

"Here's one that usually silences the room with product teams. A collaborative room planner — customers place HÖMSTYLE furniture in a floor plan, see it in 3D, design together in real time. Your product team has been asking for this for two years. Engineering scoped it at six months: WebSocket server, Redis pub/sub, Socket.io cluster, load balancer, real-time database. One Durable Object replaces all five."

"Drag the sofa. Watch the green border — you hold the lock. Nobody else can move it simultaneously. That's the DO's SQLite lock system. No Redis, no coordination service."

"Two users, two tabs. Drag in one — it moves in the other instantly. Live cursors show exactly where each person is in the room. Add a piece in Tab 2 — appears in Tab 1 immediately."

"The entire WebSocket server, the SQLite database, the pub/sub broadcast, the lock manager — it's one TypeScript class. Deploy time: under two minutes."

"One more that's directly tied to revenue — search visibility. Your product pages are a React SPA. Googlebot crawls every URL and sees the same generic title: your site name. No meta description. No structured data. Products don't appear in Google. When they do, there are no star ratings, no price — nothing to make someone click."

"A Worker intercepts the HTML, reads the product data from your existing database, and injects the right title, meta description, and JSON-LD structured data for every URL. Googlebot now sees a unique, keyword-rich title for each product. Your pages become eligible for Rich Results — stars, price, availability — directly in Google Search."

"That's what shows up in Google when someone searches for your product. Unique title, meta description, star rating, price. Zero backend changes."

"Let me build an estimate based on what you've told me about your volumes."

"This is your approximate monthly Cloudflare cost across everything we just showed. The R2 line is especially worth noting — this is what you stop paying AWS for egress. Download this as a spreadsheet for your finance team."

"Everything you've seen is deployable on your own domain today. Two scenarios here — AI chatbot on any existing site, and R2 image cache with zero egress. Both have copy-paste Worker code. The R2 one needs zero configuration — it works on whatever hostname it's deployed to."

"Your engineering team can have either of these running in under an hour."

"To summarise what we covered:"

— AI Chatbot injected via HTMLRewriter — 30-min deploy, no backend change
— R2 — $0 egress permanently, every cached asset auditable with timestamp + metadata
— Semantic search — understands intent, reduces abandonment
— Local pricing — auto-converts for any market via request.cf.country
— Subtitles — 48× cheaper than AWS Transcribe, dual-model pipeline (Nova-3 for English, Whisper for local languages)
— Voice AI — STT + LLM + TTS in one endpoint, 24/7
— Collaborative Room Designer — real-time multi-user sync, one DO replaces WebSocket server + Redis + load balancer + real-time DB
— SEO — unique title + meta description + JSON-LD Rich Results for every product URL
— Image resizing — 1.8 MB product photo → 180 KB WebP on mobile, just a URL prefix
— AI review assistant — pill tags + full review drafts, 3× review completion

None of this touched your backend. None required a sprint.

Opening question: "When your engineering team gets a new feature request — AI search, a chatbot, local pricing — what's the typical journey from approval to production? Three months? Six?"

"What you're looking at is a live demo environment. 15 use cases across AI, storage, search, caching, SEO, real-time collaboration, and more — deployed across two real websites: a furniture store and a streaming platform. Real Cloudflare APIs. Real AI models. Real data. No mocks."

"Every single feature was deployed without touching either origin server. Not a line of backend code changed. That's possible because of one concept I want to establish before anything else."

"A Worker is deployed against a route pattern — *yourdomain.com/* or more specific like *yourdomain.com/images/*. Every matching request hits the Worker before it reaches your origin."

"The Worker has full access to the request: headers, body, URL, method, and Cloudflare's free metadata — country, city, PoP, ASN. It can fetch the origin response, read it, transform it, and return something different. The origin has no idea any of this happened."

"Now I want to show you something immediately practical. Watch the bottom right corner."

"That widget was not in the HTML the server sent. The origin returned the same response it always does. The Worker piped that response through HTMLRewriter — a streaming HTML transformer built into the Workers runtime — and appended the chat widget before the closing </body> tag."

"The Worker also created a new API endpoint — POST /api/chat — that didn't exist on the origin. Worker B wraps Worker A. You can layer capabilities without merge conflicts, without touching source code."

"Now let's talk about infrastructure cost. Every image request your users make — your cloud provider charges egress. AWS S3: $0.085/GB. GCP: $0.12. Azure: $0.087. It's not visible day-to-day but it compounds."

"R2 stores the same objects, serves them from the nearest Cloudflare PoP, charges zero egress. Permanently."

"The demo ZIP here contains a built-in cache checker. Upload it to Pages as your origin, deploy the Worker, load the gallery once — then paste your Worker URL into the checker. 30 HEAD requests fire in parallel. Every card shows a green badge with the exact ISO timestamp from customMetadata, the file size, and the origin URL."

"Open the R2 dashboard right now — every object has those three metadata fields. That's not just object storage. That's a queryable audit log of when every asset entered your edge infrastructure."

"Before I show this — quick question. Do you look at star ratings on reviews? Most e-commerce teams do. The aggregate number looks fine, maybe 4.2 stars."

"Watch what the AI finds."

"Five stars. The star average includes this. But read what it says — shelf collapsed, nearly hit a child, filing a complaint. The reviewer explicitly wrote 'I am giving 5 stars so this review gets seen.'"

"distilbert-sst-2-int8 catches it immediately — NEGATIVE, high confidence. The star score said nothing. The AI saw through it."

"The Worker wraps your existing reviews API. Fetches what you already have. Runs each review through the model in parallel via Promise.all(). Returns enriched JSON with sentiment scores appended. Your reviews endpoint is unchanged."

"Inception is in this catalog. Interstellar is in this catalog. Zero results because SQL LIKE looks for the literal string."

"The Worker embeds the query using @cf/baai/bge-m3 — a multilingual embedding model that handles Thai, Indonesian, Vietnamese, and 100+ languages out of the box. Queries Vectorize for cosine similarity. Enriches from D1. Under 50ms from any PoP."

"request.cf.country, request.cf.city, request.cf.colo — free metadata on every single request. No geo-IP API subscription. No database lookup. No latency."

"This page auto-detects your location and shows local pricing. Worker reads the country, looks up the exchange rate, rewrites the price elements via HTMLRewriter. Origin sends the same HTML. User sees their currency."

"At 100,000 daily active users, you're potentially making 100,000 database calls for the same catalog data that hasn't changed in hours. KV checks the edge first. Hit: under 10ms from the PoP handling the request. Miss: query D1 once, write to KV with a 60-second TTL. The next 9,999 requests in that window never touch the database."

"AWS Transcribe: $0.024/min + S3 + IAM + batch pipeline. Workers AI Whisper: $0.0005/min — 48× cheaper. One API call."

"The pipeline: the browser decodes the audio to 16kHz mono PCM, splits into 45-second chunks with 5-second overlap, sends 3 chunks in parallel. Model routing is automatic — Nova-3 ($0.0052/min) for English, Whisper ($0.0005/min) for Thai and other ASEAN languages. Then two AI correction passes: @cf/aisingapore/gemma-sea-lion-v4-27b-it per chunk for regional language nuance, then @cf/openai/gpt-oss-120b reads the full assembled transcript, detects the video domain (medical, tech, legal), and corrects domain vocabulary globally. Two AI passes — one for language, one for domain accuracy."

"@cf/deepgram/nova-3 for STT — lower word error rate than Whisper for conversational speech, critical for a support use case where users speak naturally. @cf/meta/llama-3.1-8b-instruct-fast for LLM reasoning. @cf/deepgram/aura-2-en — arcas voice — for TTS. Three models, one Worker endpoint, under 2 seconds end-to-end."

"RealtimeKit spins up a WebRTC session via Cloudflare Calls. Visitor link for the customer, presenter link for the agent. No Zoom, no Twilio, no third-party SDK billing. The media is routed through Cloudflare — which means it works without the WebRTC latency penalty of a US-only TURN server."

"This is the Durable Objects demo. One class — WebSocket server, SQLite database, pub/sub broadcast, furniture lock manager, live cursors — all in a single TypeScript file. No Redis. No Socket.io. No load balancer."

"Three rooms, three separate DO instances. Each room's state is completely independent. env.ROOM_DESIGNER.idFromName('vardagsrum') — that's the whole routing layer."

"Green border — you hold the SQLite lock. UPDATE furniture SET x=?, y=? WHERE id=? AND locked_by=? — atomic write, broadcast delta to every connected WebSocket. Release on pointer-up."

"Every message type — grab, move, rotate, release, add, remove, cursor, reset — is handled in webSocketMessage(). One handler, no external broker."

"Three.js renders the room with recognisable multi-part furniture shapes. Click any piece to select it — raycasting hits the bounding volume, finds the group, reads the furniture ID from the mesh key. Height slider adjusts the group's Y position client-side — useful for placing a lamp on a table."

"WebSocket Hibernation: the DO sleeps when all connections are idle. Memory cost drops to zero. When the first message arrives, it wakes in under 5ms — ctx.acceptWebSocket(server) with server.serializeAttachment(session) survives the sleep cycle. 1,000 idle connections cost nothing per month."

"DO v{n} — every state change increments this counter via SQLite. It's not in memory. It survives a Worker restart. The presence strip shows every connected user's name and colour — this.ctx.getWebSockets() iterates the live sessions."

"React SPA. Every film URL sends <title>STREAMVAULT</title> to Googlebot. No meta description. No structured data. The film pages don't rank individually."

"The Worker reads the film ID from the URL, fetches metadata from D1 or KV, and uses HTMLRewriter to rewrite the <title>, inject <meta name="description">, and append a <script type="application/ld+json"> Movie schema block — all in the response stream before the browser sees it. Origin sends the same HTML it always did."

"Four API endpoints. None existed on the origin. The Worker handles them entirely at the edge — origin never receives these requests. Each is independently deployable. If one throws an error, the others keep running. No microservices orchestration, no container cluster."

Ask first: "Roughly how many API requests per month? What's your storage footprint — images, videos, files? How many AI calls per day across search, chatbot, voice?"

"Two scenarios here — AI chatbot on any existing site, and R2 image cache with zero egress. Both have copy-paste Worker code. The R2 Worker needs zero configuration — no ORIGIN to set. It adds X-R2-Bypass: 1 when fetching origin, breaking the self-loop. Deploy it to any hostname, bind an R2 bucket, done. Self-configuring."

"Two more if their engineering team wants to explore further: Image resizing — 1.8 MB product photo → 180 KB WebP on mobile, just a /cdn-cgi/image/ URL prefix, zero new infrastructure. And AI Content Generation at /demo/feedback — Workers AI generates contextual review pill tags and full draft reviews, 3× review completion rate."

Final question: "Which of these has the highest business impact for your team in the next 90 days — and what would a 2-week proof of concept look like for you?"

Stop. Let them answer.