Inspiration
In modern digital communication, we often find ourselves suffering from "Social Paralysis." When staring at a message on a dating app or a cryptic text from a friend, we struggle with two main blockers: * Decoding the Subtext: We don't know exactly what the other person means or the emotion hidden between the lines. * The "Vibe" Block: Even when we know what we want to achieve (e.g., to decline politely, to flirt, or to be assertive), we don't know how to phrase it in a way that hits the right note. While current generic LLMs can generate replies, they often fail in two ways: they are too generic/safe (giving robotic, textbook advice) and too "Old School" (failing to grasp current internet slang and trends). We built TalkSense 言感 to bridge this gap. We wanted to create not just an assistant, but a "Social War Room" that understands the nuances of modern communication.
What it does
TalkSense 言感 is a multi-agent social assistant tailored for the nuances of Chinese social media and interpersonal dynamics. It operates like a supportive "WeChat group chat" with trusted friends, each offering a distinct perspective to help you navigate complex relationships. * Deep Intent Decoding: It analyzes the input text to decipher the sender's true intent and emotional state. This is particularly crucial in Chinese communication, which is often high-context and filled with subtle subtext (e.g., distinguishing between a polite refusal and playing hard to get). * The "Squad" Response System: Instead of a single generic answer, TalkSense 言感 activates a specific team of AI Agents. They simulate a group chat discussion, offering three distinct strategies based on their unique personas: * Shuangshuang (The High-Energy Alpha / 爽爽): * Persona: "The Main Character." * Vibe: Self-centered, confident, and unbothered. * Role: She treats social interactions as an experience, not a test. She focuses on your pleasure and self-worth, helping you maintain a "High-Value" status. Her advice is about framing you as the prize, not the chaser. * Wenrong (The Empathic Ally / 温荣): * Persona: "The Ride-or-Die." * Vibe: Gentle, supportive, and emotionally intelligent. * Role: Your unconditional emotional safety net. She validates your feelings first ("I know you're upset...") before offering advice. She acts as your "mouthpiece," helping you express vulnerability or affection safely without losing dignity. * Zhangliang (The Reality Check / 张凉): * Persona: "The Stop-Loss Expert." * Vibe: Sharp, objective, and facts-over-feelings. * Role: He ignores the "fluff" and analyzes the behavioral data. If someone is wasting your time or treating you poorly, he will brutally cut through the delusion to help you "stop your losses" immediately. * Trend-Aware Generation (Chinese Context): Unlike standard models that sound like translated textbooks, our agents utilize Few-Shot Learning based on a curated, up-to-date corpus of Chinese internet trends (sourced from platforms like Xiaohongshu, Weibo, and Douyin). This ensures the replies use authentic slang, correct emojis, and culturally relevant memes that resonate with Gen Z users.
How we built it
* Dynamic Corpus Pipeline: Instead of relying on static datasets, we capture the prevailing "vibe" and trending conversational contexts from social networks (like Twitter/X, Reddit, The Red, TikTok). These cultural insights are then distilled into structured Few-Shot examples, ensuring our agents understand the nuance of current internet culture without just copying text. * Multi-Agent Architecture: We leveraged an Agent framework to assign distinct System Personas to each model. Before generating a response, each agent retrieves relevant stylistic examples (In-context Learning) to guide its tone. * Orchestration Layer: A central orchestrator manages the user input and distributes it to the agents, ensuring they respond in a conversational, "group chat" flow rather than a simple list. * Prompt Engineering: We spent significant time refining prompts to ensure the "Trendsetter" doesn't sound cringey and the "Diplomat" doesn't sound robotic.
Challenges we ran into
* The "Cringe" Factor: It is incredibly difficult to make an AI sound genuinely cool without crossing the line into "How do you do, fellow kids." Balancing the temperature and selecting the right few-shot examples was a constant trial-and-error process. * Context Nuance: Slang words (like "bet," "cap," or "gaslight") change meanings entirely based on context. Training the agents to recognize these subtleties was a major hurdle. * Latency Management: Running multiple distinctive agents simultaneously creates lag. We had to optimize with parallel processing to ensure the "group chat" felt real-time.
Accomplishments that we're proud of
- Passing the "Vibe Check": Our agents successfully generate replies that feel human and culturally relevant. The "Trendsetter" agent can actually use slang correctly! * Diverse Perspectives: The contrast between the agents (e.g., The Diplomat vs. The Savage) provides genuine value, giving users a spectrum of choices they wouldn't have thought of themselves. * Simulated Social Dynamics: We successfully built a UI that feels like a supportive group chat, offering users not just text, but emotional validation. ## What we learned
- Personality > Accuracy: In social contexts, users often prefer a response that has character over one that is grammatically perfect but boring. Strong personas are the key to engagement. * Data Expires Fast: Internet culture moves at light speed. A static dataset becomes obsolete in weeks. Dynamic injection of fresh examples is crucial. * The Need for Validation: Users don't just want a reply; they want to know their feelings are valid. Seeing an agent get angry on their behalf (The Savage) is just as therapeutic as the advice itself. ## What's next for TalkSense 言感 * Mobile Keyboard Extension: Integrating TalkSense 言感 directly into mobile keyboards for real-time suggestions while typing in apps like WhatsApp or Messenger. ## Disclaimer The content, text, and advice provided by this project are entirely generated by Artificial Intelligence (Large Language Models). These responses are simulated based on specific personas and do not represent the personal views, opinions, or values of the developers.
Log in or sign up for Devpost to join the conversation.