ShelfSense - Transforming E-commerce Through Explainable Generative Intelligence

Inspiration

We have all been there. You type a simple, human request into a search bar something like "I need wireless headphones for long work calls, good battery, not too expensive", did you ever get the answer in one go? (If yes, I cannot believe how lucky you are:). Modern e-commerce is high-choice but incredibly low-clarity. You are forced to navigate hundreds of similar products, decode unstructured and emotionally contradictory reviews, and battle keyword filters that completely miss your nuanced intent. I realized that e-commerce already has enough data; what it lacks is clarity. Shoppers are experiencing deep decision fatigue and have zero trust in black-box recommendation algorithms. I didn't want to build just another generic shopping chatbot. I wanted to build a system built for trust, not just convenience. That is how ShelfSense was born: an explainable shopping layer that bridges the gap between overwhelming product information and confident human decisions.

What it does

ShelfSense is a smart product discovery and recommendation engine. Instead of giving users another endless list of products, it creates an explainable shopping layer on top of product data. It helps users answer the questions that actually matter: What is best for me, not just most popular? and Why was this product recommended?. By connecting the shopper's true intent to structured product attributes and review evidence, ShelfSense surfaces recommendations with transparent reasoning the user can trust. It is fundamentally explainable, not just conversational.

How I built it

I built ShelfSense as a two-layer recommendation system: Qdrant for semantic retrieval and Neo4j for relationship reasoning. In Qdrant, I stored products and review chunks as vectors inside collections, with structured metadata attached as payloads, then used payload filtering, payload indexing, and hybrid dense+sparse search to retrieve highly relevant candidates quickly.

I then used Neo4j as a property graph to model products, brands, categories, features, and review themes as connected entities. Using Cypher, I enriched each retrieved candidate with graph context such as related products, shared features, and nearby evidence patterns, so the system could reason across relationships instead of treating each product as an isolated text record.

A Flask API backend orchestrated the full pipeline, and a final ranking layer combined vector relevance, metadata match, graph proximity, and review evidence before passing the result to the LLM. That helped to generate recommendations that are both relevant and explainable, which is the core value proposition of ShelfSense.

And finally the most important selling point :) built a clean React/Next.js frontend with a natural-language search box, recommendation cards, and explanation panels so a user can move from vague intent to evidence-backed choices in a single flow.

Challenges I ran into

One major challenge was dealing with review intelligence. Reviews are noisy but rich in real-world product truth, so had to build an NLP extraction pipeline to reliably pull out themes (like durability or fit) and attach sentiment and evidence to them. Additionally, pure LLM output can feel unstable for ranking, which I solved by introducing our hybrid scoring engine to add consistency.

Accomplishments that I'm proud of

I successfully built an explainable shopping layer that uses both structured and unstructured product intelligence. The MVP prototype already works on a realistic flow, taking a complex natural language query - like "I need wireless headphones for long work calls, good battery, not too expensive", and returning strong options with a short explanation for each of the requirements.

What I learned

I learned that trust is the ultimate product differentiator. Generative AI is absolutely necessary for this problem because it can interpret intent, summarize review signals, and explain recommendations in a way traditional filtering cannot.

What's next for ShelfSense

The immediate next step is preparing for the GenAI Zürich Hackathon 2026 on-site refinement and judging phase on April 1-2. Also plan to further refine UI's explanation panels to highlight review themes alongside product attributes, perfecting the "trust moment" where users see exactly why ShelfSense beats basic search.

Built With

• flask
• gemini
• hugging-face
• neo4j
• nextjs
• python
• qdrant
• rag
• react