Notes from building AI systems.

Retrieval, fine-tuning and evaluation, written up with the numbers — including the experiments that didn't work and the assumptions that turned out to be wrong.

Projects

Telugu Bhagavad Gita RAG

Grounded question answering over 699 verses in Telugu. Every quotation is verified verbatim against the source verse, and the system refuses rather than answering from the model's memory.

+81%recall@10 after fine-tuning both encoders
+83%NDCG@10, 0.256 → 0.468
100%citations verified against source
3.1sto first word, down from 45s
Live demo → Source on GitHub →

Writing

  1. My embedding model could read Telugu fine. My questions were the problem.SoonA floor test, a wrong hypothesis, and the multilingual problems that were actually real.
  2. I wrote life situations for 699 verses, and it didn't helpSoonAugmenting retrieval for thematic questions — and why the gain vanished behind a reranker.
  3. Fine-tuning BGE-M3 on $4.71 of synthetic dataSoonHard negatives, verse-level splits, and a margin that grew by pushing wrong answers down.
  4. The off-the-shelf reranker made my search worseSoonWhy a reranker only helps if it's better at the task than the retriever it reorders.
  5. 0.6% of the Ramayana has a Telugu explanationSoonBuilding a corpus in a low-resource language.
  6. Enforce, don't ask: citations, false claims, and 45 seconds down to 3SoonA prompt that changed nothing, a model swap that did, and a citation rate that silently hit zero.