02
Research
With the Ministry of Tribal Affairs, I turn welfare-scheme documents into posters and animations that low-resource language communities can actually use in a culturally faithful way, not just translated.
Retrieval · graphs
CTU-FlowRAG
Coherent Thematic Unit segmentation, then role-conditioned retrieval over typed sentence graphs. Built so a welfare document can be queried by the role the reader actually cares about.
0.82 nDCG@10 · 0.75 MRR@10 · 0.96 role coverage · 3,048 documents
Culture · RAG fusion
QCCAF
A query-conditioned fusion layer that adds local cultural context before image prompts are written. Zero-shot transfer across related low-resource languages, not just the one we trained on.
0.93 nDCG@10 on Santali · zero-shot Khasi, Bodo, Garo
Latency · bandits
Bandit-Quota
A latency-budgeted contextual bandit for per-query retriever selection. Match an ensemble when you can afford it; skip the rest when you cannot.
2.4x to 9.3x faster than full ensemble · quality match or better on 6 of 7 BEIR datasets