Hi, I'm Andrei

Built a system to solve the pain of manually writing complex JSON configurations for workflow automation tools. The friction of looking up schema definitions and debugging missing commas made me realize we needed a natural language interface: just say "send an email when task duration exceeds 2 hours" and get valid, ready-to-use JSON.

The architecture intentionally avoids heavy vector databases in favor of a lightweight TF-IDF retriever using scikit-learn. For code generation, I found that exact keyword matching for field names and enum values often yields better context than semantic similarity. The retriever identifies the three most relevant reference configs, which surround the user's prompt. Reliability is enforced by passing the LLM's output through a strict `jsonschema` validator. If the generated JSON fails validation (missing required fields or wrong types), the system catches it immediately rather than shipping broken config to production.

This project taught me that for specialized code tasks, a rigid validation loop is far more important than a smarter model. Grounding the LLM in valid examples and constraining it with a schema turned a localized problem into a reliable utility.