AgentQR

Inspiration

We hated seeing generic, blocky QR codes that ignore a brand’s personality.
Crew AI + W&B Weave sparked the idea of chaining “mini-apps” (agents) into a single, automated brand-styling pipeline.

Input: User submits a brand name.
Search: EXA Search Agent finds the URLs describing that brand’s visual identity.
Scrape: BrowserBase Scraper Agent fetches and cleans the text from those pages.
Extract: Style Extraction Agent identifies design keywords and hex color codes.
Analyze: Brand Analyst Agent turns those keywords and colors into a concise visual-identity narrative.
Prompt: Prompt Engineer Agent compresses the narrative into a comma-separated list of up to 5 keywords (≤120 characters).
Generate: QR Code Generator Agent uses that prompt to produce a high-resolution, on-brand artistic QR code.

Crew AI to orchestrate five specialized agents in a sequential pipeline.
Phi-4-mini-instruct model (within a Crew AI wrapper).
EXA.ai for semantically rich brand-identity search.
BrowserBase for reliable, REST-based web scraping.
Custom regex + LLM logic for style extraction.
W&B Weave for real-time logging of every agent call, inputs/outputs, and performance traces.

Noisy HTML scrapes: Tuning our regex to capture valid hex codes without false positives.
Prompt Engineering: Crafting prompts that consistently generated high-quality, cute, and artistic QR code images was non-trivial. It required iterative tuning and explicit instructions to the LLM to get the desired style and quality.
Ensuring the frontend correctly parsed and displayed the backend’s nested response structure (especially for the QR code image URL) was a subtle but critical bug that required careful debugging.

The Power of Multi-Agent Systems: Orchestrating specialized agents (research, scraping, style extraction, prompt engineering, art generation) can produce richer, more context-aware outputs than a monolithic approach.
Prompt Engineering is Key: The quality of AI-generated art is highly sensitive to prompt wording. Explicit, example-driven, and style-focused prompts yield much better results.

Optimize the Pipeline: Reduce end-to-end latency and improve aesthetic quality by fine-tuning agent parameters and parallelizing non-dependent steps.
Expand Agent Ecosystem: Add specialized agents (e.g., typography extractor, layout analyzer) to surface deeper style cues and support niche design languages.
Interactive UI & CLI: Build a lightweight web interface and command-line tool so non-technical users can generate branded QR codes with a few clicks or commands.