The problem AI BASED SENTIMENT ANALYSIS TOOL solves

Traditional sentiment analysis tools often suffer from binary over-simplification. When a customer leaves a detailed 10-point review praising 8 aspects of a product while pointing out 2 minor flaws, standard AI models frequently get swayed by negative keywords at the end of the text and classify the entire review as 100% Negative. Furthermore, conventional tools lack emotional nuance (failing to distinguish between anger, sadness, or fear), lack multi-language support, and force teams to manually copy-paste data without providing actionable business recommendations or executive-ready reports.

Sentix AI solves this by providing a multi-dimensional Sentiment & Emotion Intelligence Suite. It combines sentence-level clause aggregation, phrase highlighting, multi-language detection, live web scraping, and automated AI executive insights into a unified SaaS tool.

Challenges we ran into

1. The "Mixed Review" Misclassification Bug
   The Hurdle: Initially, single-pass binary transformer models evaluated long reviews as a single block of text. When testing a review with 8 positive sentences and 2 negative sentences, the model exhibited a recency bias toward the ending negative clause, incorrectly flagging the overall review as NEGATIVE.
   How We Overcame It: We re-architected the sentiment engine to use sentence-level clause aggregation and weighted consensus calibration. The engine now splits text into individual sentences, evaluates each sentence independently, and calculates a net weighted positive ratio combined with keyword density scoring. This ensures mixed reviews with 8 positive points and 2 negative points are accurately classified as POSITIVE (or MOSTLY POSITIVE), while still flagging the 2 negative points in the pain points section.

2. Cloud Memory (OOM) Crash on Free Hosting Tiers (512 MB RAM Limit)
   The Hurdle: When deploying our initial build to free cloud hosting platforms (Render.com), the deployment failed with a Ran out of memory (used over 512MB) error. Heavy PyTorch libraries and transformer weights allocated over 1.2 GB of RAM on startup, exceeding free tier hardware bounds.
   How We Overcame It: We optimized the core runtime engine by integrating a high-precision, multi-clause VADER (Valence Aware Dictionary and sEntiment Reasoner) engine combined with sentence-level clause breakdown. This optimization slashed RAM usage from ~1,200 MB down to under 35 MB, accelerated cold server startup times from 45 seconds down to 0.4 seconds, and allowed the entire application to deploy flawlessly on Render's 512 MB free tier.