Bhavika Kadlecha

I’m an AI & ML engineer with an MSc in Computer Science (AI & ML), and I build LLM applications that hold up outside of demos. Most of my work sits in what I call the reliability layer: strict schemas for model outputs, deterministic validation, fallbacks for when APIs fail, and benchmarks that show whether a change actually helped. I also write things up clearly, with documented trade-offs and results, so a remote team can pick up my work without needing a call.

What I’ve built:

• Hybrid-RAG corporate policy retrieval engine: Uses Docling to extract structured content from PDFs, including complex tables. Combines ChromaDB dense and BM25 sparse indexing with Reciprocal Rank Fusion and cross-encoder re-ranking. In a retrieval ablation, MRR@5 improved from 0.708 (dense-only) to 0.875 (full pipeline). Containerized with Docker.

• LLM-as-a-Judge evaluation pipeline: Benchmarks RAG outputs for factual consistency and context relevance using strict Pydantic schemas. Runs 12 concurrent evaluation cases in under 3 seconds with pytest-asyncio and the Groq API. SQLite prompt-hash caching removes redundant re-evaluations, and custom pytest hooks generate HTML dashboards and JSONL benchmark reports.

• Multimodal invoice extraction and vendor reconciliation pipeline: An async FastAPI microservice that uses multimodal LLMs via LiteLLM to extract line items from invoice PDFs and images into validated Pydantic models. Deterministic Python checks verify the arithmetic, with bounded schema-repair retries. Lanczos downscaling reduced image payload size by over 60%. ChromaDB lookup flags unverified or mismatched suppliers.

• Learnovo (IIT Jodhpur internship): A full-stack learning platform (Django, React, SQLite) that turns free-text learning goals into structured study plans using the Groq API (LLaMA-3) with strict JSON schemas. A rule-based fallback serves pre-computed curricula during API timeouts, and React localStorage keeps learner progress between sessions.