AI Review Intelligence Platform
E-commerce & Grocery Brands
Measured Outcomes
72%
Exact star match
97%
Within ±1 star
83%
Sentiment accuracy
Business Challenge
Consumer brands and marketplace sellers face the same bottleneck from three angles: ops teams can't read every review, support needs fast routing for negative feedback, and product teams need structured insight from review archives — not generic summaries. Off-the-shelf models hallucinate star ratings, output inconsistent labels, and can't plug into refund queues, catalog fixes, or competitive research workflows.
Engineering Solution
An end-to-end review intelligence system: a fine-tuned Qwen2.5-14B model for exact star prediction and sentiment, a Gradio demo for instant single-review analysis, batch pipelines for catalog-wide triage, and aggregation workflows that surface themes, listing gaps, and competitive insights.
Implementation Process
- 1
Data curation & stratified training
Filtered 568K Amazon Fine Food Reviews down to 50K high-quality, deduplicated samples with balanced 1–5★ distribution so the model learns every rating band — not just five-star praise.
- 2
QLoRA fine-tuning & evaluation
Fine-tuned Qwen2.5-14B-Instruct with 4-bit LoRA SFT, assistant-only loss, and W&B monitoring. Built a held-out eval suite with per-star accuracy, confusion matrices, and best-checkpoint selection.
- 3
Live demo & batch inference
Shipped a Gradio web app and CLI for instant single-review analysis, plus batch handlers that ingest CSV/JSONL and return structured JSON with predicted_star, sentiment, and key observations.
- 4
Triage routing & insight aggregation
Added rules-based routing (e.g. Negative + ≤2★ → urgent queue) and aggregation workflows that roll up star distributions, sentiment mix, and recurring themes for listing optimization and competitive research.
Business Impact
One fine-tuned model powers three production-ready workflows: instant star and sentiment prediction from any review text, automated ops triage that surfaces high-risk feedback first, and product intelligence reports that turn review archives into actionable themes — all running on a single consumer GPU (~10GB VRAM) with 100% structured output parse rate for downstream automation.
