
NeuroDesk AI Agent is a production-ready, multi-agent digital marketing platform built on AMD Developer Cloud using the MI300X GPU and vLLM inference engine. It is designed to autonomously plan, execute, and optimize full marketing campaigns with minimal human input. The system is built around one Master Agent Brain (MAB) that orchestrates three specialized Sub-Agent Brains (SABs): SAB1 — Audience Intelligence & Personalization Engine: Analyzes user records across CRM, web analytics, and purchase history. Uses K-means clustering to identify micro-segments and high-LTV cohorts. Generates personalized email sequences and dynamic ad creatives per segment. Delivered 41% higher open rates and 28% CTR increase in targeted cohorts. SAB2 — Creative Orchestrator & Multimodal Generator: Produces 1,350+ on-brand assets across 12 channels using context-aware prompt engineering. Auto-generates responsive video ads for TikTok, Reels, and YouTube. Reduced asset production time from 8 hours to just 45 minutes with 98% brand compliance. SAB3 — Real-Time Optimization & Feedback Loop: Deploys a reinforcement learning model that analyzes campaign performance every 15 minutes. Dynamically adjusts bids, rotates creatives, and pauses underperforming variants. Achieved +23% conversion rate and -17% CPA vs. baseline. The entire inference stack runs on AMD MI300X (192GB VRAM, 20 vCPU, 240GB RAM) via vLLM 0.17.1 on ROCm 7.2, giving the system the raw compute power needed to run large language models like Qwen2.5-7B in real time. The agents communicate in a closed feedback loop — SAB3's performance data refines SAB1's segmentation, and SAB1's insights shape SAB2's creative briefs — making the system self-improving over each campaign cycle. Overall Campaign Results: CTR improved from 2.1% → 3.8% (+81%), ROAS improved from 3.1x → 4.5x (+45%), and CPA dropped from $48 → $39.84 (-17%). Built by team Anuj (Apdevlab) for the AMD x lablab.ai Hackathon 2026.
10 May 2026