yenupam

Blog & Projects

i’ve been breaking and building things since gpt-3.
ai, linux, old hardware. if it can be hacked, i’m interested.
sometimes i train models. sometimes i write about it.

Recent Projects

  • Matra: A sentence-constrained superword tokenizer for Indic scripts. On the MUTANT eval set: aggregate fertility 1.78, 72.5% sequence reduction, 9.10 bytes per token. Produces 31% fewer tokens than Gemini 3.5 Flash, 62% fewer than Qwen-3.6-MoE, and 22% fewer than Sarvam-105B across 24 languages.
  • Drut (WIP): An edge Speech-to-Speech model targeting sub-150ms latency, multilingual (first of its kind), full-duplex interactivity, and native tool calling.
  • Desi Maximalism: Text-to-Image LoRA made from a hand-curated dataset to flawlessly generate the vibrant, nostalgic retro aesthetics of Desi Indian culture that frontier models fail to reproduce.

Latest Posts