Research

Testing whether structural constraints, not prompts, can make AI systems reliable enough to trust with regulated decisions.

Research topics:

bharath@research: ~

$ ls research/

$ cat graph-constrained-retrieval/README.md

Naive RAG retrieves by similarity alone and misses evidence that's structurally adjacent to the question. My MSc dissertation constrains retrieval with graph structure instead, FAISS ANN search re-scored by NetworkX two-hop expansion, tested across three pipeline variants over 1,585 papers and 46,882 chunks.

$ cat graph-constrained-retrieval/metrics.log

hallucination_reduction: 94.5%

faithfulness: 0.083 -> 0.95

provenance_coverage: 47.76% -> 95%

context_precision: 0.50 -> 0.92

context_recall: 0.55 -> 0.93

answer_relevancy: 0.65 -> 0.95

abstention_unsupported: 98%