I Built WanderWise: A Local-First AI Planner That Wants You to Close the Screen

Most AI products compete for more screen time. The architect of WanderWise sought to create a system whose definition of success is getting closed. WanderWise AI takes a brief sentence, such as "I need a calm reset after work and I have one hour," and turns it into a concise outdoor plan. It recommends an activity, explains the match, breaks up the available time into manageable steps, lists what to bring, and provides one relevant safety reminder. Then it encourages the user to close the phone.

The architect recognized that the simple advice "Go outside" rarely translates into actionable steps. When someone is tired, restless, bored, or short on time, choosing an activity becomes another planning task. Search results exacerbate this issue, providing dozens of generic listicles, route pages, and recommendations that lead to even more screen time before the user ever steps outside.

To solve this problem, WanderWise requires only five inputs from the user: the kind of outdoor experience desired, how much time is available, who will be joining, the current weather, and the level of energy. The system then generates one tailored plan, rather than another generic feed.

The current version of WanderWise includes twelve curated outdoor activities, ranging from a slow "noticing walk" for a quiet reset to a micro-hike for higher-energy users. Each activity provides structured information on suitable weather, duration, energy requirements, company size, equipment needed, safety considerations, and a three-part sequence of actions. The result is intentionally compact, allowing users to copy the plan, close the browser, and head outside.

WanderWise utilizes an open-source AI model, Apache-2.0-licensed all-MiniLM-L6-v2 through Transformers.js. This model runs in the browser, converting the user's natural language input and activity descriptions into 384-dimensional embeddings. Cosine similarity is then used to determine which activity is most semantically aligned with the user's needs.

While semantic similarity is the primary ranking signal, WanderWise combines it with explicit constraints to ensure practicality. The final score is calculated as a weighted average of semantic similarity (62%) and compatibility factors for duration, weather, energy, and company. This hybrid ranking system leverages the model's strength in understanding intent while relying on conventional code to handle logistical considerations.

Running the open model locally offered several benefits to the project. It kept user prompts private, eliminated the need for API keys and metered inference billing, and allowed for full transparency into the activity data, ranking weights, and model choice. The static application has no user database, analytics SDK, or AI backend, and the model is downloaded and cached upon first use to ensure offline functionality.

During live deployment, a bug was discovered where the model was pointed to an incorrect ONNX repository. However, a transparent fallback mechanism ensured that the system still provided a useful activity suggestion. This bug led to improvements in model loading, clearer labeling of fallback recommendations, and versioned service-worker assets for improved reliability.

The design of WanderWise is intentionally minimalist, avoiding the chatbot format that often invites additional questions and prolongs engagement. Instead, the system presents a single form and recommendation card containing all essential information for the outdoor activity. The visual style incorporates natural elements like forest green, warm cream, leaf green, and trail orange to evoke the feeling of an upcoming outing rather than a productivity tool.

The architect tested the system with a focused recommendation check using the prompt "just," demonstrating WanderWise's ability to generate a suitable plan without excessive user input.