WhatShouldIBuy - Consumer Decision Support
A consumer decision-support web app for high-stakes purchases (laptops, phones, TVs, ACs, mattresses, headphones). Built end-to-end from product spec to deployed live app. Designed to act as a decision assistant, not a product filter.
What it is
A web-based consumer decision-support application that helps users make confident, transparent purchasing decisions for high-stakes products. Built end-to-end as an independent project: product spec, UX, and deployed live app.
Built with: Replit (app), ChatGPT (spec), Claude (verification), Gamma (presentation).
The problem
Consumers face information overload, biased reviews, unclear personal preferences, and rapidly changing product options when making high-value purchases. Most tools filter and rank. None explain.
The goal: act as a decision assistant, not a product filter. Go beyond showing options to explaining the reasoning behind each recommendation.
Key features
- Structured preference collection via a progressive questionnaire (1-2 questions per screen, skippable, with progress bar). Designed explicitly to reduce cognitive overload and improve completion rates.
- Conversational search with text and voice input. Users type natural language queries. The system interprets intent, asks clarifying questions if needed, and triggers the recommendation engine.
- Decision Engine with five modules: Scoring Model, Confidence Model, Tradeoff Analyzer, Regret Risk Analyzer, Explanation Generator.
- Weighted scoring: Budget Match (30%), Feature Match (25%), Performance (20%), Reliability (15%), Popularity (10%).
- Decision Confidence Score: a calculated percentage based on preference depth, match strength, and variability among top options.
- Regret Risk Indicator surfaces potential future concerns to address long-term satisfaction, not just immediate fit.
- Tradeoff Insights explicitly identify what each recommendation is best at. A Why Not These? transparency layer explains why popular products were excluded.
- Side-by-side comparison view with clickable recommendation cards linking directly to purchase pages.
What I learned
Recommendations are easy. Confidence calibration is hard. Telling a user why we excluded the obvious answer is what makes a recommendation system feel like an advisor instead of a black box.