Cinematic Sentiment Engine

💡 Inspiration (The Problem)

In the film industry, audience feedback is everything. But when I tested standard, out-of-the-box AI sentiment analyzers, I noticed a massive flaw: they completely fail to understand human nuance.

If an audience member writes, "The visual effects were absolutely mind-blowing, but the systemic corruption subplot felt heavy-handed," a standard NLP model will flag it as 99.9% NEGATIVE simply because it sees words like "corruption" and "heavy-handed." It entirely misses the praise. I realized that binary (Positive/Negative) classification is practically useless for real-world studio executives who need to understand complex, mixed audience reactions.

⚙️ What it does (The Solution)

The Cinematic Sentiment Engine is a context-aware AI dashboard built for the entertainment industry. Instead of basic binary classification, it utilizes a highly advanced 28-class emotion detection model.

When a user submits a review, the engine:

1. Identifies the Dominant Emotion: (e.g., Admiration, Amusement, Disappointment).
2. Provides a Nuance Breakdown: It extracts the top three conflicting emotions to show the layers of the audience's reaction.
3. Generates Studio Recommendations: It translates raw mathematical probabilities into actionable business logic (e.g., advising a studio to "Highlight quote in marketing materials" or "Flag pacing issues for the writers' room").

🛠️ How I built it

• Frontend: I built a commercial-grade, responsive UI using Streamlit, featuring a two-column layout, simulated processing telemetry, and dynamic color-coded metric bars.
• Machine Learning Backend: I leveraged the Hugging Face Transformers library, specifically implementing the SamLowe/roberta-base-go_emotions pipeline.
• The "Judge-Breaker" Logic: By manipulating the top_k=3 parameters of the pipeline, I forced the model to return a multi-dimensional array of emotions, which I then parsed into a custom business-logic engine.

🚧 Challenges I ran into

The biggest hurdle was overcoming the limitations of pre-trained binary classifiers. My initial prototype successfully ran a standard DistilBERT model, but the results were too flat to be useful. I had to pivot my entire backend strategy mid-hackathon to integrate a multi-label RoBERTa architecture that could actually comprehend contextual subtext without breaking my frontend UI. I also overcame local environment configurations and Git deployment hurdles to get the project version-controlled and live!

🏆 Accomplishments that I'm proud of

I am incredibly proud of transforming a raw Python script into a fully realized, polished web app entirely on my own. Adding the "Studio Recommendation" feature bridged the gap between pure machine learning math and real-world business value. It doesn't just show data; it makes decisions.

🚀 What's next for the Cinematic Sentiment Engine

In the future, I want to scale this application to scrape live Twitter/X feeds and Rotten Tomatoes pages via API. This would allow a studio to monitor the "Nuance Breakdown" of thousands of live reactions on opening night in real-time, giving them an unprecedented dashboard for audience retention.

Built With

• huggingface-transformers
• machine-learning
• python
• pytorch
• streamlit