Consistent Pictogram Generation for AAC

I’ll demo how I built an AI-powered pictogram workflow that combines three components:

1. An offline prompt-to-image matching architecture that maps user intent to the most relevant pictograms (generated or existing), including how I represent prompts, store image/text embeddings locally, and select candidates with lightweight retrieval—so the experience stays fast and consistent.

2. Pictogram generation using an image generation model (prompt → pictogram), including the constraints and post-processing steps I use to maintain a consistent pictogram style and usable output format.

3. A lightweight “AI builder” workflow that assembles pictograms into simple compositions, focusing on the interaction model and the data structures that make the builder feel like a practical tool rather than a one-off generator.

The demo will walk through each workflow step-by-step, paired with technical implementation notes and lessons learned around consistency, latency, and reliability.