-
-
cover
Inspiration
Everyone faces conversations they dread—asking for a raise, delivering difficult feedback, or navigating a personal conflict. The fear of saying the wrong thing often leads to avoidance, anxiety, and missed opportunities. We realized that while pilots have flight simulators and athletes have practice drills, there was no safe, accessible space for the average person to practice and master these critical communication skills. Our inspiration was to build that "flight simulator"—a personal gym where anyone could rehearse challenging dialogues, build confidence, and learn to navigate any conversation with skill and grace.
What it does
ConvoCraft is an AI-powered role-playing platform that transforms your browser into a personal dialogue coach. Here’s how it works: Define Your Scenario: You set the stage by describing the context of the conversation and defining the AI's persona with a core goal (e.g., "to reject a deadline extension") and a personality style (e.g., "supportive but firm"). Engage in Dynamic Dialogue: The AI generates a detailed, realistic persona and initiates a conversation. Powered by Google Gemini, the AI responds in real-time, staying perfectly in character and adapting to your every word. Practice Offline: For quick practice or without an internet connection, ConvoCraft offers an offline mode with local AI templates and a powerful "paraphrase" tool (🎲) to add variety to any text on the fly. Receive Instant Feedback: After your conversation, the AI provides a comprehensive performance review, analyzing your strengths and areas for improvement across communication skills like assertiveness, empathy, and strategic approach.
How we built it
ConvoCraft was built with a focus on accessibility, performance, and a powerful hybrid AI model, all using vanilla web technologies. Frontend: The entire application is a responsive Single-Page Application (SPA) built with pure HTML, CSS, and modern JavaScript (ES Modules). We avoided heavy frameworks to ensure a lightweight and fast user experience. Primary AI Engine: For high-fidelity dialogue, persona generation, and performance analysis, we integrated the Google Gemini API (gemini-1.5-flash-latest), which serves as the core "brain" of the simulation. Local AI Engine (Offline Functionality): We used Transformers.js (@xenova/transformers) to run a quantized flan-t5-small model directly in the browser. This powers the instant paraphrasing feature and offline templates, ensuring privacy and accessibility for all users. User Experience: All interactivity is handled with vanilla JavaScript DOM manipulation. The user interface features CSS variables for easy theming, and the Web Speech API provides immersive voice synthesis for both the user and the AI persona.
Challenges we ran into
AI Prompt Engineering: Getting the Gemini-powered persona to be both realistic and concise was a major challenge. Early versions were too verbose. We solved this through iterative prompt refinement, adding specific instructions like "keep your response dense and to the point" to create more natural, impactful dialogue. Managing the Hybrid AI Model: Integrating both a server-side API and a client-side model required careful state management to ensure a seamless user experience, clearly indicating which features were available with and without an API key. Cross-Browser Voice Consistency: The Web Speech API behaves differently across browsers. We had to implement a robust voice selection logic with fallbacks to ensure the chosen persona (e.g., a male manager) would have a fitting voice whenever possible.
Accomplishments that we're proud of
The Hybrid AI Architecture: We are incredibly proud of creating a tool that works powerfully both online and offline. The combination of Gemini for depth and Transformers.js for instant, private utility is a key accomplishment that makes ConvoCraft uniquely versatile. The Complete Feedback Loop: ConvoCraft isn't just a chat simulation; it's a true learning tool. The "Request Feedback" feature closes the loop, turning practice into measurable improvement by providing users with actionable insights. Infinite Replayability: By allowing users to define any persona and scenario from scratch, we've created a tool that can be adapted to virtually any conversational challenge, making it endlessly useful for personal and professional growth.
What we learned
Throughout this project, we learned three key lessons. First, the art of prompt engineering is paramount; minor changes in instruction can radically alter AI behavior and quality. Second, providing offline-first features, like the local AI rewriter, delivers immense value and builds user trust. Finally, we reaffirmed that with a solid architectural approach, vanilla JavaScript remains an incredibly powerful tool for building complex, modern web applications without the overhead of large frameworks.
What's next for ConvoCraft
The future for ConvoCraft is focused on making it an even more immersive and intelligent coaching tool. Advanced Analytics & Progress Tracking: We plan to implement a dashboard where users can track their performance metrics (e.g., assertiveness, empathy) over time and across different scenarios. Voice Input: To make the simulation even more realistic, we will integrate the Web Speech Recognition API, allowing users to speak their responses instead of typing. Community Scenario Library: We envision a feature where users can save and share their most effective scenario configurations, creating a community-driven library of practice modules. ConvoCraft Pro (SMZ Edition): A future premium version could offer access to more advanced language models, team-based training features, and even deeper psychological analysis of conversational styles.
Log in or sign up for Devpost to join the conversation.