Published Paper Details:

Today, amid the ever-evolving financial markets, tools like Chartink play a vital role in finding investment targets. Nevertheless, their dependence on manually constructed filters and quasi-programmingbased syntax results in serious usability issues. To tackle these problems, we propose a next-generation cloudbased solution based on the dual Large Language Model (LLM) architecture that will let you easily scan the financial market using simple and conversational NLP input. The proposed system uses an innovative approach to filtering data dubbed as Programmatic Filter Synthesis. In this process, the Qwen3-14B model from the OpenRouter API creates a JSON object including SQL query parameters that are validated by our system and safely executed against the DuckDB analytical engine running queries on 5.5 GB of Apache Parquet data. Moreover, our system is equipped with a local guard module known as Qwen3Guard-Gen-0.6B that filters out all the possible hazardous or inappropriate inputs before processing them via the generation module. The entire stack includes the following: web front-end: React/Next.js 16; server-side: Python FastAPI with SSE streaming, caching: Redis service and Supabase authentication. We achieved an 87% adjusted semantic accuracy rate while testing our system on 100 benchmark prompts.