The Problem: When Professional Networking Becomes a Habit Loop
LinkedIn has evolved from a career tool into a highly optimized doomscrolling engine and I fell right into its engagement trap. To break the cycle, I wanted to design a silent, "always-on" spatial agent that could subtly observe my screen and deliver a contextual nudge the moment it detected aimless scrolling.
The core user experience was designed to be beautifully simple: input a behavioral boundary, and let the background LLM handle the vigilance. However, moving from a static concept to an invisible, seamless agent required an intense, highly iterative technical execution.
The technical challenge
Running continuous screen analysis via cloud-based APIs gets expensive fast. To make this practical, I shifted to a local-first architecture using LM Studio to power the agent via an open-source model (Gemma 4 12B QAT).
While my initial setup handled the front-end orchestration incredibly well, running a local model in real-time introduced immediate performance bottlenecks. Optimizing the experience required getting under the hood to balance AI intelligence with machine hardware:
Prompt Engineering: Architected a precise system prompt to guide the agent's evaluation cadence, ensuring it could accurately distinguish actual professional research from passive doomscrolling.
Latency vs. Performance: Aggressively dialed down the model's temperature to suppress "agent thinking" noise, forcing faster, more deterministic interventions.
Hardware Guardrails: Fine-tuned the GPU thread allocation to keep the local LLM from resource-starving the system and crashing my workspace.
Reflections
While the experiment was a technical success—the agent successfully detected and flagged my scrolling habits, the real-world UX hit a classic hardware limitation. Running a continuous vision-and-text pipeline locally introduced noticeable lag during heavy design workflows, making an "always-on" local state impractical for daily production.
This project proved to be an incredible masterclass in the realities of local-first AI design. It highlighted the delicate trade-offs between user privacy, operational compute costs, and the threshold of acceptable device performance.