I build RAG systems that survive production.

I'm Yamraj Khadka, a computer engineer based in Kathmandu, Nepal. I specialize in retrieval-augmented generation, domain-specific LLM fine-tuning, and the evaluation work that makes both of them trustworthy.

What I actually do

Most AI projects fail between the demo and production. The model is rarely the problem. The problem is that the right evidence never reaches the model, nobody defined what "correct" means, and the system has no way to say "I don't know."

I work on the unglamorous parts:

• Document ingestion for PDFs, scans, tables, and inconsistent formats
• Structure-aware chunking with metadata that enables filtering
• Hybrid retrieval (BM25 + dense) with reranking
• Grounded generation with citations and refusal behavior
• Evaluation harnesses: recall@k, MRR, faithfulness, latency, cost

Flagship work: E-Wakil

E-Wakil is a legal AI assistant for Nepali law. It answers questions over statutes and case law with citations, built on a fine-tuned Mistral-7B and a hybrid retrieval pipeline. Fine-tuning taught the model legal phrasing; retrieval supplied the actual law. Facts never went into the weights.

Try the live Nepal Law AI demo →

Open source

I maintain the RAG Eval Kit, a dependency-free retrieval evaluation library with BM25, dense, and hybrid retrievers, plus recall, precision, nDCG, and MRR metrics. It exists because most teams measure RAG by vibes.

How I work

Fixed-scope engagements with a measurable acceptance test. Every project starts with a paid diagnostic so we both know what "working" means before code is written. You get the code, the evaluation set, and documentation you can run without me.

Stack

Python, FastAPI, TypeScript, PyTorch, Hugging Face, PEFT/LoRA, LangChain/LlamaIndex, pgvector, Qdrant, Mistral, Llama, OpenAI, Docker, Modal/RunPod.

Contact

Email: hello@yamrajkhadka.com.np
Or book a 20-minute diagnostic.