← Back to projects

Smart Trust Meter

BS Computer Science Thesis · 2024

An AI-powered dealership trust intelligence platform for car buyers in Lahore, Pakistan. It aggregates real Honda and Suzuki reviews, scores them with DistilBERT multilingual sentiment (English + Roman Urdu), computes an overall trust score, and breaks trust into eight service aspects so buyers can search, compare dealers, and submit new feedback with instant AI analysis.

Live DemoResearch Project

The challenge

When researching a dealership, buyers had to rely on Google Maps: open one searched place and skim a pile of generic star ratings and unstructured reviews. There was no aspect-based place to judge pricing, service quality, parts availability, repair quality, communication, and more. Comparing dealers meant guessing from noisy overall scores, especially when reviews mixed English and Roman Urdu.

Business requirements

  • Focus on Lahore auto dealerships (Honda + Suzuki) with real customer review data
  • Ingest Google Maps-style review exports and persist dealerships + reviews in MongoDB
  • Score each review with multilingual sentiment covering English and Roman Urdu
  • Compute an overall trust score per dealership from positive vs negative sentiment
  • Expose aspect-based trust across 8 dimensions: pricing, service quality, parts availability, time efficiency, repair quality, cleanliness, expertise, communication
  • Let users search dealerships, open detail pages with trends/aspects/recent reviews
  • Support side-by-side compare for 2–4 dealerships
  • Allow users to submit new feedback and get instant AI sentiment analysis
  • Show trust score trends over time (monthly + cumulative)
  • Provide admin import paths for dealership CSVs and aspect payloads
  • Deploy a usable demo stack (frontend + API + ML) beyond a local notebook

The approach

I built Smart Trust Meter end to end: React trust dashboard, Express API, Flask ML service with DistilBERT (distilbert-base-multilingual-cased) as the active model, aspect-labeled dealership data, search/compare/feedback flows, and cloud deploy on Vercel + Render + Hugging Face Spaces with MongoDB Atlas.

What was done

  1. 01

    Business data & aspects

    Imported Google Maps-format review CSVs for 10 Lahore Honda/Suzuki dealerships, labeled sentiment and eight service aspects offline, and stored dealership trust objects (score, mentions, positivePct) for product use.

  2. 02

    NLP model pipeline

    Trained and evaluated DistilBERT multilingual against a legacy dual-input LSTM baseline, with Roman Urdu augmentation lifting RU accuracy from ~46% to ~79% on held-out eval (English ~90%, F1 ~0.92).

  3. 03

    Product requirements delivery

    Shipped home search, dealership list/detail, aspect cards, 2–4 dealer compare, live feedback scoring with ML + lexicon fallback, trust-trend charts, and an AI model metrics page.

  4. 04

    Ops & deployment

    Containerized ML + API + frontend for local Docker demos; deployed production path across Vercel, Render, Hugging Face Spaces, and MongoDB Atlas with CI checks for frontend, backend, and NLP.

Results

  • Aspect-based trust scores across 8 service dimensions per dealership
  • Active DistilBERT multilingual model (English ~90% acc, Roman Urdu ~79% after augmentation)
  • Search, compare (2–4), feedback, and trust-trend flows for Lahore dealers
  • Deployed across Hugging Face Spaces, Render, and Vercel

Technologies

PythonDistilBERTNLPReactExpressMongoDBHugging FaceRenderVercel