Hauska AI

The problem

Design projects fail on vision misalignment before the first mockup. Stakeholders describe outcomes in words; designers interpret in pixels; the gap between those two is a month of revision cycles and rework that nobody scoped for. The traditional response has been "more meetings, more docs, more moodboards," which trades one form of friction for another without actually closing the interpretation gap. Hauska's thesis is that AI-generated visual alternatives, produced live in a shared canvas, change the shape of the conversation — from "let me try to describe what I mean" to "which of these three is closer to what you want." The client picks; the designer refines; iteration happens on pixels instead of on paragraphs.

Key challenges

Real-time collaborative canvases are hard on their own. Add AI-generated visual alternatives (with latency, cost, and quality variance) and you have three difficult product problems interacting: real-time sync (WebSocket + conflict-free state), AI generation pipeline (queuing, prompt engineering, fallback), and multi-participant UX (whose selection wins, how iteration branches). Each is a full engineering discipline; combining them in one shared surface is where most of the work actually goes.

What we built

AR Data delivered engineering on Hauska's AI-assisted design canvas — real-time visualization from prompts, side-by-side alternative generation, and shared iteration state across clients and designers. AI turns intent into visual options so the conversation happens on pixels, not paragraphs. Architecture: WebSocket-based real-time sync with CRDT semantics for conflict resolution; generation pipeline that batches AI requests, caches near-duplicates, and streams intermediate previews; shared iteration state that tracks branches per participant and merges when consensus emerges. The canvas surfaces the AI as a peer participant — visible in the collaborator list, transparent about what it's generating — rather than a hidden layer.

Our approach

1. 1

   AI as visible peer participant

   The AI shows up in the canvas as a labeled collaborator rather than as invisible magic. This design choice made the tool feel honest to designers, who otherwise found the invisible-AI framing uncomfortable.

2. 2

   Streaming intermediate previews

   Instead of waiting for final images, the pipeline streams intermediate previews so clients see the direction quickly. Perceived latency drops even when actual latency is fixed.

3. 3

   CRDT-based real-time sync

   Multi-participant canvases need conflict-free state; CRDTs make that tractable without a central server arbitrating every keystroke.

4. 4

   Near-duplicate caching in the generation pipeline

   Design iterations often re-request small variants. Caching near-duplicates cuts AI cost meaningfully while preserving the impression of unbounded generation.

Key architectural decisions

Results

• Faster vision alignment between clients and designers
• AI-generated visual alternatives in real time
• Real-time collaborative canvas with CRDT-backed sync
• Reduced rework cycles by front-loading interpretation to the shared canvas
• Streaming previews improve perceived latency without added infrastructure cost
• Near-duplicate caching keeps AI cost in a viable range for professional use
• AI visible as collaborator, not hidden — designers stayed in control

Impact

Hauska represents a class of AI-in-collaboration product where the AI's value is not "do the work" but "make the conversation about the work faster." The engagement produced patterns we now bring to any collaborative-AI product: AI as visible peer, streaming intermediates, CRDT-based state, and cache-heavy generation pipelines. The designers who use it stay in the driver's seat; the AI just makes the client's intent legible faster.

Tech stack