Work / PolyAI

Dialog RSN-1

Launch identity for a new reasoning model: a dark developer palette, generative particle imagery, a directed film and 4K frames.

rsn1-loop.mp4
Role
Set the identity, directed the film shoot, built the generative imagery system, the model frames, the launch page design and the ad sets.
With
Ideal Insight (film) · PolyAI research and web teams
Scope
Launch, Identity, Film, Generative
Year
2026
Agents
  • The particle imagery is generated by a shader I built, not drawn frame by frame. Agents helped port it between the launch page, the film inserts and the ad sizes.

Dialog RSN-1 is PolyAI's new reasoning model. It launches on 30 September 2026 with its own page, a film and a paid campaign.

The brief

A model isn't a product with a screen. It has a name, some numbers and a claim about how it behaves on a call. The job was to give it an identity that sits under the PolyAI brand, speaks to developers and enterprise buyers in the dark, technical register they respond to, and stretches from a 4K launch frame to a 300 pixel ad without falling apart.

What I did

Set the identity: a dark palette, a generative particle system for all of the imagery, and the type and layout for the launch page. Directed the film shoot with Ideal Insight, a table, a laptop and the people who built the model, so the launch has real people showing real behaviour rather than a voiceover over stock. Built the model frames at 4K and the ad sets that went to demand generation.

Model frame, excerpt. Every frame in the launch comes from the same particle system.rsn1-loop.mp4

The imagery

The particle field isn't a still I made once. It's a WebGL2 shader that renders any image or video as glyph particles, so the launch page, the film inserts and the ads all come from one system at whatever size they need. I wanted this model to look different from the rest of the range: the same system, tuned to read as circular where the others read as fields. How the shader works is written up in the lab.

An annotated transcript: a customer's line, the detected sentiment, then the agent's reply
An annotated conversation. Latency, turn-taking and sentiment shown on the transcript instead of described.rsn1-ad-conversation.webp
The benchmark frame: 3x faster than GPT Realtime, with two numbers underneath
The benchmark frame at 4:5, the size that works hardest on LinkedIn.rsn1-ad-benchmark.webp

Where it went

Launch page, film, model frames, social and paid. The paid sets are in Demand generation.