AI agent

Kage

Live

Kage

AI agent·Live

WHAT IT IS

Most chat agents on websites are forgettable. Generic avatars, “How may I assist you today?” tone, trained mostly to deflect. They sit in the corner like a receptionist who can’t actually book the meeting.

Kage is our attempt at the opposite. A chat agent with a voice, a job, and a clear sense of when to send you somewhere else. It lives on the homepage as proof. We say we build AI agents, so here’s one. It also works as a filter that pre-shapes inquiries before they hit the contact form.

The fact that it’s here on Labs is the point. We built it for ourselves first.

HOW IT WORKS

Most chat agents optimize for engagement, keeping you talking as long as possible. Kage optimizes for qualification. Figuring out who you are and pointing you somewhere useful, fast. Three paths cover almost every visitor.

THE INTERFACE

Most of the work isn’t in the model. It’s in the surface. Page-aware starter chips so the conversation begins differently on /projects than on /about. A thinking-state avatar that closes its eyes while a response streams in, then opens them when it lands. A mobile experience that takes over the full screen instead of squeezing into a corner. The character has to feel like it belongs to the brand before anything it says will land.

STACK

Claude APIEdge RuntimeUpstash RedisStreamingNext.js API Routes

THE BUILD

We built Kage in an afternoon. The avatar took less than an hour. A few iterations through Nano Banana Pro, background removed, shipped. The brain is Claude. Streaming runs on Edge Runtime so it works on Vercel’s Hobby plan. Upstash Redis handles rate limiting and session caps, because anything pointed at the open internet needs a floor.

What’s interesting is what we left out. No retrieval layer, no fine-tuning, no vector database. Not because those tools don’t have a place, but because Kage didn’t need them. The job is small and well-defined. A sharp system prompt with the studio’s actual knowledge baked in does most of the work. Real projects, real clients, real positioning. The instinct to over-engineer the AI layer is one of the more expensive habits in this space right now.

WHAT’S NEXT

A few things on the bench. RAG for project-specific deep-dives, so Kage can speak with detail about specific case studies instead of summarizing. Conversation analytics for knowing which questions actually get asked. High-intent alerts that ping us when a real lead is in the chat.

Each one earns its place or it doesn’t get added. The version live right now is intentionally simple, and we’d rather keep it that way than ship features for the sake of motion.

Try Kage

Kage is in the bottom right of every page. Click the avatar, ask anything. The conversation is real, the agent is the same one we just walked you through.

Generative

Growing Minds

Live

Growing
Minds

GenerativeLive

Studio shots, lifestyle imagery, video, and a live storefront for a product that didn’t physically exist yet.

THE BRIEF

We were already building the Growing Minds Shopify store when the need for product creative came up. Studio hero shots, lifestyle scenes, video, an animated storefront. Premium positioning. Scientifically credible without feeling clinical.

No physical product existed yet. No jar to photograph. No gummies to arrange on a counter. Just a label file and a manufacturer’s spec sheet.

We’d been deep in generative AI tooling on our own time and pitched the approach: give us the files, we have the pipeline and the compute. Let’s see what comes out.

Clear plastic filled with small translucent gummies is essentially a worst-case scenario for AI image generation. Transparency through transparency. A label with real text that has to stay legible across every shot. Details that drift between frames if you’re not careful. That’s what made it interesting.

Each layer of the problem got its own tool.

THE GUMMY

Before you can fill a jar, you need something to fill it with.

A gummy sounds simple. Small dome of colored gelatin with sugar coating. But getting AI to make one that looks real, not like a plastic bead or a gumdrop from a stock photo, took way more iteration than expected.

Wrong color. Too smooth.

Closer. Sugar still off.

That’s the one.

THE JAR

The manufacturer’s spec sheet had dimensions, wall thickness, thread count for the cap. Enough to model from scratch, but we wanted to get closer faster.

We found the manufacturer’s website had a 3D model of the jar embedded on a product page. We used an LLM to scrape the file, pulled it into Blender, and had a dimensionally accurate starting point within hours instead of days. From there we refined the geometry to exact spec: PET plastic with correct refraction, 0.7mm wall thickness, IOR 1.2.

Ask any AI model to generate a jar and it’ll happily give you one. But the proportions drift between shots. Cap slightly different height. Shoulder taper changes. For a product launch where every image has to feel like the same physical object, that inconsistency kills trust immediately.

One 3D model. Same geometry fed into every tool downstream. The anchor for the whole project.

THE BUILD

Four views of the same geometry. Wireframe, mesh, solid, and final material.

The Blender file became the single source of truth. Every render, every AI generation, every video frame traced back to this model. Change one dimension here and it cascades through the entire pipeline.

Same jar. Every tool downstream.

PUTTING IT TOGETHER

AI’s not great at text. Some models are better than others, but the small print is where they all struggle. The ingredient list, supplement facts, certification badges. The stuff that has to be letter-perfect or it’s useless.

We were given the label as a flat design file. Getting it onto the jar meant wrapping it around a curved surface in Blender and locking it there. Every render tool we tried wanted to reinterpret the text. Garbled characters, warped type, hallucinated side panels. The fix was to never let AI touch the label at all. Designed in 2D, mapped onto 3D geometry, baked into every reference image downstream. AI handled lighting, materials, and scene composition. The words stayed ours.

Filling the jar was its own problem. Gummies don’t stack neatly. They tumble in and settle wherever gravity puts them. Some lean against the wall. Some pile up in the center. AI doesn’t understand any of that. Early attempts gave us gummies that floated, clipped through the walls, or arranged themselves in perfect rows. The physics were wrong in a way that’s hard to explain but immediately obvious.

We ran Blender’s rigid body simulation to model how gummies would actually land inside the container. Getting the sugar dusting right took its own round of iteration. Too uniform and it looks sprayed on. Too sparse and the gummy looks wet. The final version has the uneven crystal scatter you’d see on a real gummy pulled from the bag.

THE ASSEMBLY

Label locked. Gummies settled. One jar.

Label locked in Blender.

+

Physics simulation for natural settling.

The final assembly.

THE PIPELINE

No single AI tool could have produced this. The project worked because each layer got the right one.

BlenderNano Banana ProKling 3.0ComfyUIGSAP + Shopify

Blender for structural truth and physics simulation. Nano Banana Pro for photorealistic rendering. Kling 3.0 for video. ComfyUI with ControlNet for the hybrid work that needed both 3D precision and AI style. GSAP and Shopify for the live animated storefront.

3D is the identity anchor.

When consistency across dozens of shots matters, route the structural identity through 3D geometry. Use AI for rendering, not for shape.

Prompt economy beats prompt complexity.

Past a certain threshold, more detail causes models to discard inputs and improvise. The approved baseline almost always outperformed the “improved” version. When iterating, subtract.

AI video scale is controlled by images, not words.

The start frame is the only reliable lever for subject distance. Prompt-based spatial language is largely ignored.