Enterprise Policy & Claims Assistant

Challenges The biggest challenge was combining several AI components into a reliable, usable application and presenting their results clearly. Making answers traceable The AI needed to base its answers on the uploaded documents. I linked answers to specific passages and added checks for source references so users can see where a statement comes from. The accuracy of the information still needs to be verified. Implementing document search without a paid embedding API I switched to locally computed embeddings, which represent text numerically so the system can find relevant passages based on meaning. This required ensuring that the embedding model was compatible with the stored document index. Handling errors and API limits Usage limits and failed model responses made the LLM integration more challenging. I added Groq as a provider, separated configuration from application code, and implemented clear error messages for missing API keys, exceeded limits, and invalid responses. Making quality measurable I used Ragas to build an evaluation pipeline that assesses answers against a prepared set of reference questions. I also had to handle interrupted test runs correctly so that incomplete results would not be treated as a passed quality check. Connecting the technology with the user experience PDF uploads, document search, and AI-generated answers needed to work together across the frontend and backend. This required clear loading states, error handling, and a reproducible setup using Docker. The ML model’s score also needed to be presented in a way that made its role in supporting manual review clear. This taught me that an AI application needs more than a working model integration. Users need to be able to verify results, understand their limitations, and know how to proceed when something goes wrong.