Hot Takes Only

Inspiration Scrolling through social media, I realized everyone’s chasing the perfect post, but no one’s brave enough to tell the truth. What if your AI could be that brutally honest best friend who roasts you into influencer greatness? That’s how Hot Takes Only was born, a self-improving AI agent that turns your selfies into data-driven glow-ups, one savage roast at a time.

What it does Upload a selfie, get roasted, glow up. Hot Takes Only uses an AI vision model to analyze your photo’s pose, lighting, and vibe, then delivers a voice-based roast that is both funny and constructive. It pulls in current selfie trends from the web, learns from feedback using Weights & Biases, and keeps evolving its sense of humor and advice style so its takes stay sharp and current.

Users receive a spoken roast through VAPI, a social appeal rating, improvement tips, and trend-based suggestions on what’s working right now. Behind the scenes, the agent continuously refines its approach based on logged performance and user engagement.

How I built it I used Next.js, TypeScript, and TailwindCSS for the frontend to create a fast and simple user experience. The backend was built with FastAPI in Python, which handled image uploads, roast generation, and Weights & Biases logging. OpenAI’s GPT-4o-mini handled vision analysis and roast creation, while VAPI generated the voice output. Browserbase was used to scrape selfie trends to influence the roast tone. The self-improvement logic was tracked and simulated through Weights & Biases to show the agent evolving over time.

Challenges I ran into Tuning the roast tone was difficult, it had to be honest but still fun. Integrating the voice layer quickly while ensuring stability under time constraints was another challenge. Handling image uploads and API latency in a short timeframe required careful design. Simulating self-improvement meaningfully in a 24-hour hackathon was also tricky, but I made it work with creative metrics and mock data.

Accomplishments that we’re proud of I built a functional, self-improving AI roast coach in 24 hours. The demo is engaging, funny, and memorable. I successfully integrated voice, vision, and trend awareness into a cohesive product. The self-improvement loop in Weights & Biases gave our agent a real sense of evolution.

What I learned I learned how to design AI interactions that balance humor and helpfulness. I saw how Weights & Biases can help visualize and track agent behavior over time. I learned to work across multiple APIs quickly and efficiently. Most of all, I learned that a well-executed mix of comedy and AI is an incredibly powerful way to engage users.

What’s next for Hot Takes Only I plan to expand the app to handle short videos and TikToks for roast and editing feedback. I want to add social feedback loops, where users upvote the funniest or most accurate roasts to train the agent. I also want to introduce themed roast personalities, such as Gordon Ramsay mode or Gen Z influencer mode. Longer-term, I hope to implement real reinforcement learning to improve based on genuine engagement metrics and eventually release a public beta version.

Built With

• browserbase
• fastapi
• next.js
• openai-gpt-4o-mini
• python
• tailwindcss
• typescript
• vapi
• weights