WellLoop AI: Predict and Prevent Team Burnout

💡 Inspiration

In the high-pressure world of software development, burnout isn't just a personal feeling—it's a systemic failure. We noticed a recurring pattern: Engineers drowning in context switching, Product Managers misaligned on priorities, and Designers feeling disconnected from the roadmap.

Existing tools operate on lag indicators—surveys that tell you a team is unhappy after they've already started quitting. We wanted to build something predictive. We asked ourselves: What if an AI could detect the subtle signals of burnout 24 hours before they become a crisis?

Thus, WellLoop AI was born for Snowfest 2025.

🏗️ How We Built It

We built WellLoop AI as a hybrid application, leveraging the speed of a modern web framework with the analytical power of Python for heavy lifting.

The Stack

Frontend: Next.js 14 and Tailwind CSS for a responsive, accessible dashboard.
AI Core: Google Gemini 2.5 Flash for rapid inference and summarization.
RAG Engine: A custom Python module using google.generativeai embeddings.

The Intelligence Layer

The heart of our system is the WellLoopContextEngine. We moved beyond simple keyword matching to Semantic Search. We use the Gemini models/embedding-001 to vectorize team sentiment.

To match a current blocker (query) with historical solutions (documents), we calculate the Cosine Similarity between their vector embeddings.

$$ \text{similarity}(\mathbf{A}, \mathbf{B}) = \frac{\mathbf{A} \cdot \mathbf{B}}{|\mathbf{A}| |\mathbf{B}|} = \frac{\sum_{i=1}^{n} A_i B_i}{\sqrt{\sum_{i=1}^{n} A_i^2} \sqrt{\sum_{i=1}^{n} B_i^2}} $$

Where:

$ \mathbf{A} $ is the vector for the user's current query (e.g., "Engineering fatigue").
$ \mathbf{B} $ is the vector for a historical document in our knowledge base.

This math allows us to mathematically quantify "relevance" and surface insights that a simple keyword search would miss.

🧩 Challenges We Faced

1. The "Cold Start" Problem

One of our biggest challenges was handling the "cold start" problem for new teams. Without historical data, the RAG engine had nothing to retrieve. Solution: We implemented a "seed" database of anonymized organizational psychology patterns. This gave the AI a baseline of wisdom to draw from on Day 1, which improves as the team adds their own data.

2. Hybrid Environment Complexity

Connecting a Next.js Edge execution environment with a Python logic layer presented latency hurdles. Solution: We optimized our architecture to keep the "hot path" (UI interactions) purely in Node.js, calling out to the Python context engine only for deep-dive analysis requests.

🚀 Accomplishments & What We Learned

Gemini's Versatility: We learned that Gemini isn't just a chat bot; its embedding models (task_type="retrieval_document") are incredibly powerful for classification tasks.
Mathematical Intuition: Implementing the vector math from scratch gave us a deeper appreciation for how high-dimensional space works in AI.

🔮 What's Next for WellLoop AI

We plan to introduce Manager Auto-Pilot—an agent that doesn't just report burnout, but automatically drafts calendar adjustments (like "No-Meeting Wednesdays") for approval when risk levels spike.