Have you seen those new AI-generated review summaries on Amazon? They are incredibly useful for buyers, but there’s a catch: they are completely locked inside Amazon’s ecosystem.
If you are a developer, PM, or data scientist trying to analyze 5,000 scattered App Store reviews, Shopify comments, or Zendesk tickets, you are still stuck doing it manually or relying on basic word clouds.
I wanted to fix that. So, I built NEXUS 🧠—a production-grade Review Intelligence Engine that brings that exact "Amazon-style" AI analysis to any dataset.
Here is a deep dive into the architecture and how I put it together. 👇
🏗️ 1. The Deep Learning Baseline
Before jumping into massive pre-trained models, I wanted to establish a strong, custom baseline.
The Data: Trained on the Sentiment140 dataset (1.6 Million records).
The Architecture: I built a custom deep Bidirectional LSTM using TensorFlow/Keras. I utilized a 128-dim Embedding layer and stacked Bi-LSTMs to capture deep contextual sequences.
Optimization: Used aggressive Dropout(0.5) layers and EarlyStopping on validation loss to halt training dynamically and restore the best weights, preventing overfitting.
🤖 2. The Transformer Inference Pipelines
To achieve zero-shot classification and granular emotional analysis in the live app, I loaded lightweight HuggingFace pipelines directly into memory:
Sentiment: DeBERTa-v3 for highly accurate Zero-Shot classification (Positive, Neutral, Negative).
Emotional Topography: RoBERTa-go_emotions to extract 28 micro-emotions, which I mapped to heuristic scores (Joy, Frustration, Urgency, Resolve).
⚙️ 3. The "Amazon-Style" Intelligence Engine
Here was the biggest challenge: heavy generative LLMs (like DistilBART) consume massive RAM and are prone to hallucination.
Instead of relying purely on an LLM to write the summary, I wrote a deterministic Component-Impact Engine. It uses Regex and Pandas to chunk sentences, extract hardware/software components (battery, screen, software, ports), calculate the failure/praise rates of each, and dynamically synthesize a natural language summary.
The output? Exactly what engineering needs to see: "Customers heavily praise the screen and UI, but express deep frustration with the battery life."
✨ 4. The Frontend UX/UI
Streamlit is fantastic for Python devs, but out-of-the-box, it can look a bit generic. I wanted a premium, glossy feel. I injected hundreds of lines of custom CSS to override the default DOM, creating a "glassmorphism" aesthetic with animated micro-interactions, gradient borders, and custom Plotly charts.
NEXUS doesn't just say a review is "negative"—it tells the engineering team exactly what is breaking so they can push a fix faster.
I'd love to hear your thoughts! Have you experimented with DeBERTa vs. custom Bi-LSTMs for your own sentiment projects? Let's chat in the comments! 💬
Link- https://sentimentanalyser-ucccl9ut869ugpmqid2ttg.streamlit.app/
Top comments (2)
Strong execution on this—especially the decision to move from pure generative summarization to a deterministic component-impact engine. That hybrid approach is exactly what most production NLP systems end up converging toward when latency, cost, and hallucination risk become constraints.
From a systems perspective, there’s a clear opportunity to take this further through collaboration on a few dimensions:
We’ve been working on multi-source review intelligence pipelines (App Store + Zendesk + CRM feedback) with a focus on real-time ingestion and vector-based clustering for theme drift detection. Your sentiment + emotion stack (DeBERTa + RoBERTa-go_emotions) would pair well with a streaming layer where embeddings are continuously updated and summaries are recalculated incrementally instead of batch mode.
Would be interesting to explore a joint extension:
plug your Component-Impact Engine into a real-time event pipeline (Kafka / Redis Streams)
add semantic dedup + anomaly detection on review bursts
benchmark LSTM vs transformer features under low-data domain adaptation scenarios
If you’re open to it, happy to align on a small POC around dataset fusion + live dashboard intelligence layer.
Sorry for the delayed response—I've been completely tied up with academics lately and just saw this! Thank you so much for the kind words and for taking the time to share such thoughtful feedback.
I completely agree with your assessment. Moving to the deterministic Component-Impact engine was a necessary step for managing those latency and hallucination constraints in production.
Your multi-source review pipeline sounds like exactly the right environment to take this to the next level. I’ve been looking for ways to push the DeBERTa + RoBERTa-go_emotions stack into a true real-time streaming context, so plugging it into your Kafka/Redis setup makes perfect sense.
I'm fully on board with the roadmap you outlined:
Integrating the engine directly into the live event pipeline.
Tackling the semantic dedup and burst anomaly detection.
Running those LSTM vs. transformer benchmarks under low-data scenarios.
Let’s absolutely align on a small POC around dataset fusion and building out that live dashboard intelligence layer. I love pushing these kinds of technical boundaries in my projects, so this collaboration is a great fit.
Let me know when you have some time to sync up and sketch out the initial architecture. You can contact me through my email, as I'll mostly be available there!