Inspiration
I come from Rajasthan, a state known for its rich heritage, culture, and tourism. Tourists from across India and around the world visit Rajasthan, but many are unfamiliar with local prices and often end up paying significantly more than the actual market rates.
Having lived in different states before returning to my home state, I noticed how unfamiliarity with local pricing makes tourists vulnerable to overcharging. I experienced a similar issue while studying in Bhubaneswar, where language barriers often make communication with local service providers, particularly auto-rickshaw drivers, difficult.
This made me think: What if tourists could know whether a quoted price is fair before making a payment?
That idea led to VeriFee, a platform designed to help travellers identify unfair pricing through community-verified data, machine learning, NLP, and AI.
What it does
VeriFee is an AI-powered tourist price-protection platform that helps travellers identify potentially overpriced products and services before paying.
Its key features include:
- Fair Price Estimation: Compares quoted prices against relevant local price ranges to help tourists understand what they should reasonably pay.
- Scam Detection: Uses statistical analysis and machine learning to identify unusually high prices and flag potentially unfair transactions.
- Community-Powered Reports: Allows tourists to upload receipts and price reports, building a shared database of real-world pricing experiences.
- Intelligent Report Retrieval: Retrieves relevant reports for a particular product, service, and location to provide evidence-based price assessments.
- City-Wise Price Comparison: Helps travellers compare reported prices across different cities.
- Multilingual Translation: Helps overcome language barriers between tourists and local service providers.
- AI Bargaining Coach: Provides contextual bargaining suggestions and helps tourists negotiate more confidently.
- Pre-Payment Alerts: Flags suspicious or unusually high prices before users complete a transaction.
The platform also considers user-selected criteria and retrieves relevant receipt information to make price assessments more contextual.
The goal is not simply to label every expensive transaction as a scam, but to provide tourists with useful evidence and context so they can make informed decisions.
How we built it
I built VeriFee by combining traditional machine learning and statistical techniques with modern LLM-based capabilities.
Technology Stack:
- Frontend: Next.js, React.js
- Backend: Node.js, Express.js
- Database: MongoDB
- ML Microservice: Python, FastAPI, scikit-learn
- AI: Groq LLM APIs
- Retrieval and NLP: TF-IDF, cosine similarity, RAG, NER
Machine Learning Pipeline:
I implemented an on-server ML pipeline combining:
- TF-IDF and cosine similarity for retrieving relevant pricing reports.
- Z-score and IQR-based anomaly detection for identifying unusually high prices.
- Linear regression for price forecasting and analysing pricing trends.
Community-Powered RAG System:
When users upload receipts and price reports, the information is stored in MongoDB along with relevant contextual details.
When another tourist checks a price, the system retrieves relevant reports based on the product, service, location, and other contextual attributes. This allows the system to ground its responses in actual community-submitted evidence rather than depending entirely on an LLM's general knowledge.
Hybrid AI and Algorithmic Decision System:
One of the most important architectural decisions was separating statistical and algorithmic decisions from LLM-generated assessments.
When at least three relevant community reports are available, the system uses algorithmic analysis to evaluate the price. When sufficient reports are unavailable, it falls back to LLM-based contextual assessment. This helps distinguish evidence-backed analysis from AI-generated estimates.
NLP and Conversational Intelligence:
I developed a seven-intent NLP pipeline incorporating Named Entity Recognition, session-aware context, anaphora resolution, and Hindi-English query normalization. This enables natural conversations, contextual follow-up questions, translation, and bargaining assistance.
Challenges we ran into
One of the biggest challenges was determining whether a transaction was genuinely overpriced or simply different because of location, product variations, or service conditions.
Some of the major challenges included:
Relevant Report Retrieval: Retrieving the correct community-submitted reports for a particular product, service, and location was difficult. Irrelevant reports could lead to misleading price assessments.
Algorithm vs. LLM Decisions: I needed to distinguish situations where sufficient community evidence existed from situations where the system had to rely on LLM-based estimates. Designing the three-report threshold and keeping these decision paths separate required careful consideration.
Dynamic Price Ranges: As new reports are uploaded, reference price ranges need to evolve. Ensuring that new reports influence subsequent assessments efficiently was an important engineering challenge.
Bargaining Logic: Designing the bargaining coach required more than generating conversational responses. I had to determine when bargaining should begin, when it should stop, and what a reasonable negotiation range should be based on retrieved pricing evidence.
AI and Backend Integration: Integrating ML algorithms, LLM services, database retrieval, and conversational NLP into a unified application required clear separation of responsibilities across the frontend, backend, and ML microservice.
Scalability and Performance: Understanding caching, database retrieval optimization, service separation, and handling growing community-generated data required additional engineering effort.
Accomplishments that we're proud of
- Built and deployed a complete full-stack tourist price-protection platform.
- Implemented an ML pipeline using TF-IDF, cosine similarity, Z-score, IQR, and linear regression without depending on external ML APIs for core price-analysis workloads.
- Developed a community-powered RAG system with provenance tracking to distinguish verified receipt evidence from AI-generated estimates.
- Designed a hybrid decision-making architecture combining statistical analysis with LLM-based contextual assessment.
- Developed a seven-intent NLP pipeline with entity recognition, contextual conversation handling, and Hindi-English query normalization.
- Integrated multiple AI capabilities into a single application using a modular backend architecture.
Most importantly, I am proud that VeriFee addresses a problem I have personally observed. It is an attempt to use technology not just to provide information, but to help people make better decisions in unfamiliar places.
What we learned
Building VeriFee helped me understand how traditional machine learning techniques can complement modern LLM-based applications.
I gained deeper practical experience with TF-IDF, cosine similarity, statistical anomaly detection, NLP, RAG architectures, and contextual information retrieval.
I also learned that simply integrating an LLM does not automatically make an application reliable. The quality of retrieved information, decision boundaries, data provenance, and fallback mechanisms are equally important.
On the engineering side, I gained a better understanding of caching, database retrieval optimization, API design, service separation, and building scalable AI-powered web applications.
One of my biggest takeaways was learning how to translate a real-world problem into a combination of statistical models, retrieval systems, conversational AI, and practical product features.
What's next for VeriFee
VeriFee is an ongoing project, and there are several improvements I would like to introduce:
- Category-Specific Price Intelligence: Develop dedicated pricing and anomaly-detection capabilities for different product and service categories.
- Live Location Awareness: Use a tourist's location to retrieve more relevant nearby pricing information.
- Live Local Marketplaces: Integrate real-time market information to improve price comparisons.
- Auto-Rickshaw Fare Intelligence: Develop specialised fare estimation and bargaining capabilities for local transportation.
- Improved Community Verification: Strengthen receipt validation, report quality, and confidence scoring.
- More Regional Languages: Expand translation and conversational support to additional Indian languages.
- Improved Retrieval and Scalability: Further optimise RAG, caching, and database retrieval as community reports grow.
My long-term vision is to make VeriFee a trusted travel companion that helps tourists navigate unfamiliar markets, understand local prices, communicate across language barriers, and avoid being unfairly overcharged.
VeriFee – Travel with confidence. Know the price before you pay.
Built With
- anomaly-detection
- artificial-intelligence
- cosine-similarity
- express.js
- fastapi
- groq
- javascript
- linear-regression
- llms
- machine-learning
- mongodb
- named-entity-recognition
- natural-language-processing
- next.js
- node.js
- oauth
- python
- rag
- react.js
- rest-apis
- retrieval-augmented-generation
- scikit-learn
- tailwind-css
- tf-idf
