SE Talk Track

This is the demo to show right after Live Object Detection, because the contrast is the whole point. That demo returns structured output — labels and coordinates, from a fixed or open vocabulary. This demo returns reasoning — you ask a real question in plain English and get a real answer, in the same conversational shape as a chatbot. Same underlying idea (Workers AI sees pixels), completely different output contract: structured data for automation vs. natural language for a human. And because Llama 4 Scout is natively multimodal, this isn't a separate vision model bolted onto a text model — one endpoint, one model, one bill.

Demo Steps

1 Upload a photo (or use the camera) — anything works: a product, a room, a document, a piece of equipment
2 Ask something genuinely open-ended: "what is this?", "does this look damaged?", "what would you guess this costs?"
3 Ask a FOLLOW-UP question about the same photo — point out this is multi-turn, the model remembers the conversation
4 Contrast with Live Object Detection: that demo gives you boxes and labels; this one gives you an answer a human could act on
5 Call out realistic use cases live: accessibility (describing a scene for a blind or low-vision user), retail ("does this outfit work together"), field inspection ("is this part damaged"), document triage ("what kind of form is this, and what do I do with it")
6 Be upfront about the one real limitation: this is a vision-language model giving its best-guess answer, not a certified inspection tool — flag that explicitly if the audience brings up safety or compliance use cases

Cloudflare Products

Workers AI@cf/meta/llama-4-scout-17b-16e-instruct

"How is this different from ChatGPT vision?" — Functionally similar in what it can answer, but architecturally different in where it runs: this calls a Workers AI model directly from a Cloudflare Worker at the edge, with no third-party API key, no separate vendor bill, and the same account/billing/observability surface as every other Cloudflare product you already use. It's also natively multimodal in one model call — no separate OCR/vision step bolted onto a text model.

🔎 Accessibility / Retail / Field Service

Visual Q&A Assistant

Point a camera or upload a photo, then ask any question about it in plain English. Open-ended visual reasoning, not structured detection — no fixed classes, no bounding boxes, just a conversational answer, with follow-ups about the same photo.

The Problem

""Object detection tells us WHERE things are, but our users ask open-ended questions — 'does this outfit match', 'is this part damaged', 'what kind of form is this' — that don't fit a fixed class list or a bounding box.""

The Outcome

natively multimodal Workers AI model, zero separate vision pipeline. One chat-style endpoint answers any question about any photo, with follow-ups.

Live demo below

Productionising this

What changes when you ship this for real

Rate limiting

This is a much more expensive call per-request than object detection (larger model, chat-style pricing). Add a KV-backed rate limiter per session/IP before this ships publicly — a single motivated user asking 50 rapid-fire questions costs real money with no natural throttle like a polling interval.

Image size & upload limits

Uploaded files are resized client-side to a max 1024px dimension before sending, but a malicious client can bypass the frontend and POST an arbitrarily large payload directly to the API. Enforce a hard server-side max body size / max decoded image dimension, and reject anything larger with a clear 413 rather than letting it hit the model.

Multi-turn token cost tradeoff

Follow-up questions in this demo intentionally do NOT resend the image — only the first turn includes it, and later turns rely on the model's own prior answers as context. That is materially cheaper (no repeated image tokens per follow-up) but means a follow-up asking about a detail nobody has mentioned yet can be answered less reliably, since the model is reasoning "from memory" of its own earlier answer instead of looking at the pixels again. If your use case needs every follow-up to be as visually grounded as the first question, resend the image on every turn instead and budget for the token cost accordingly.

Hallucination risk — read before using for anything with real consequences

This model gives its best-guess natural-language answer — it is not a certified inspection or safety tool, and it can be confidently wrong. Never use unreviewed vision-model output for pass/fail decisions with real consequences (structural safety inspections, medical assessments, compliance sign-offs, accessibility descriptions where an incorrect answer creates real risk). Any production use case in that category needs a human in the loop reviewing the model's answer, not an automated action taken directly on it.

Bandwidth

A single 800×600 JPEG capture (or a resized upload) is typically 60-150KB — negligible compared to the continuous polling loop in Live Object Detection, since this is one photo per session rather than a frame every 1.2 seconds.

Content moderation

If this is exposed publicly, add an image moderation step (or Workers AI's own moderation-capable models) before passing arbitrary user-uploaded photos to the vision model, especially if answers or photos are ever logged or displayed to other users.