Quartz

An AI-powered Chrome extension that protects sensitive data in prompts and optimizes LLM calls by masking private info, rerouting queries, and crafting clearer, cost-efficient responses.

Comment 1

Inspiration

In the present day and age, AI tools are becoming essential and many are growing with AI being an integral part of their lives. Unfortunately, every prompt/question risks exposing sensitive data such as names, client info, financial details, API keys, etc. At the same time, individuals and teams waste tokens and time on inefficient, unclear prompts that increase costs, and reduce LLM performance (unsatisfactory answers).

What it does

Introducing Quartz, an all-in-one Chrome extension that sanitizes your data and enhances your prompts for better and efficient responses.

  • Quartz detects sensitive information in your prompt and alerts you to it allowing you to replace it with anonymized information
  • Quartz re-routes conversational and ambiguous queries to a smaller model which uses lesser resources
  • Quartz engineers your prompts to get clear and high-quality answers

How we built it

  • Security/sensitive info detection: A natural language processing model (NER) that scans your prompt and determines which parts contain sensitive information and allows for replacement with a single click of a button. The model runs on huggingface inference to reduce latency (important).
  • Resource Saving: The user prompt is evaluated for its complexity and ambiguity to determine the need for a response which is provided by the small language model (meta-llama 17b parameter models - very low resource consumption in comparison to GPT, Claude, etc.)
  • Prompt Enhancement: Prompts are re-engineered using a research-backed 5C framework and the Gemini-2.5-flash-lite which ensures that the output is high-quality and covers all aspects mentioned

Challenges we ran into

  • Deploying the functions on cloudflare workers was tough at the beginning as it was our first time
  • Diverse testing to determine the viability and availability of NER models (for sensitive info identification) on gpu inference
  • Integration of all tools to ensure low latency and positive results with GenAI models

Accomplishments that we're proud of

  • Proud of being able to deploy most of the services to the cloud and building a chrome extension for the first time
  • Being able to pivot with technologies when faced with challenges especially with regards to identification of sensitive information

What we learned

  • Planning is necessary, having the big picture from the beginning ensured that there were no internal conflicts with tasks
  • Having a team we enjoy to work with, motivated us to keep going until the very end of the hackathon
  • Practicing pitching is very important and is an essential part of the hackathon, we enjoyed building the pitch and would love for you to see it

What's next for Quartz

  • Pushing into the tech space as a Saas for integration into enterprise solutions (we want companies to adopt AI safely and effectively and be a part of that journey)
  • Building our own NLP models covering a larger context window, faster inference and covering more types of sensitive information

Built With

Share this project:

Updates