Therapy Beagle
Resisting Overwhelming AI Through Comfort and Care
TherapyBeagle challenges the hyper-efficiency trend in AI design by offering a slow, warm, and human-centered alternative. Built entirely locally through LLaMA 3 (via Ollama), Beagle invites users to breathe, pause, and reflect — not just respond.
Client
Self
Services
Product Design UI/UX Human Computer Interaction
Industries
Human-Computer Interaction (HCI)
Date
April 2025
TherapyBeagle is a locally hosted, emotionally intelligent chatbot that responds with care, warmth, and reflection. Designed as a response to the sterile, high-efficiency feel of most AI tools, it poses a simple question: What if AI could feel like a kind, supportive dog,rather than a robotic assistant? The project addresses the loneliness and emotional fatigue users may feel when interacting with impersonal technology. Instead of transactional, fast-paced interactions, TherapyBeagle slows things down and encourages introspection. Target Users: 1. Individuals seeking non-judgmental emotional support 2. Users wary of data privacy in commercial chatbots 3. Students and creatives navigating stress or isolation Technical Implementation Technology Stack: 1. LLM backend: LLaMA 3.2 (via Ollama) — hosted locally 2. Web framework: Flask (Python) Frontend: HTML, CSS, and JavaScript (theme toggle, accessibility) 3.Memory system: Custom SessionMemory class using JSON file persistence Key Features: 1. Contextual memory: Remembers last 6 turns of conversation Emotion detection: Tags themes like "grief," "loss," or "anxiety". 2. Dark/light theme toggle Typing animation dots for “thinking” feedback Architecture Overview: 1. main.py: Routes POST requests from the user input, builds dynamic prompts, queries the LLM, stores memory 2. SessionMemory: Manages user state, profile, emotional tags, and conversation logs 3. index.html: Renders the chat UI, feedback bubbles, and reset mechanism
User Feedback and Iterations Testing and Insights: 1. Users described Beagle as “gentle,” “non-judgmental,” and “better than Replika” 2. Many users opened up more freely knowing the bot wasn’t cloud-hosted 3. Key request: memory! Users wanted Beagle to recall what they had said earlier What Changed Based on Feedback: 1. Added persistent memory using a local sessions/ directory 2. Introduced emotional keyword detection for grief, stress, and reflection 3. Tweaked prompt structure to reflect emotional continuity (“Last time, you mentioned…”) Outcomes and Impact Quantitative Outcomes (local use): Average conversation length during peer testing: 7+ turns 1. 70% of users said they’d use it again “when feeling overwhelmed or stuck” Qualitative Learnings: 1. People respond deeply to a chatbot that doesn’t try to fix them — just one that listens 2.Running locally gave users a sense of privacy and control not found in mainstream tools 3. Emotional design isn’t just a UI feature — it must be baked into tone, pace, and memory Reflection: "TherapyBeagle taught me that compassion can be coded — not through logic, but through pacing, warmth, and memory."