Projects on yenupam

Drut (WIP)

Tue, 16 Jun 2026 00:00:00 +0000

Status: Work in Progress. This is an active design exploration. Specifications and targets are subject to change.

System Architecture

Interactive diagram: open in Excalidraw

System Arch.

The goal is a speech to speech model that runs entirely on a phone. Not a cloud API with a thin client. The model should live on device, listening continuously, ready to respond in under 150 milliseconds. It should call tools, handle interruptions mid sentence, and work across languages. Here is the current design direction.

Matra

Sun, 07 Jun 2026 00:00:00 +0000

Technical report: matra_paper_200k.pdf

LLM with Matra as tokenizer is still pending. If anyone wants to help me with compute, hmu @ e-mail

Live Demo

note: visit huggingface space link if the embed fails to load.

Matra: A Tokenizer Built for India’s Languages

200k, 128k vocab size. 23 languages. The lowest token counts and highest compression across the board.

In active development, more optimzation, tests, language specific tokenizers coming soon.

Most tokenizers treat Indian languages as an afterthought, fragmented syllables and half-formed characters scattered across a vocabulary designed for English. Matra was built from the ground up to fix that. The result, benchmarked on June 15 2026 against the IN22-Gen test split at 200k tokens per language, is unambiguous: Matra produces fewer tokens, denser representations, and dramatically lower fragmentation than GPT-5, Gemini 3.5 Flash, Gemma-4-31B, Qwen-3.6-MoE, Sarvam-105B, and Sutra-v2 across all 23 scheduled languages of India.

Desi Maximalism

Mon, 01 Jun 2026 00:00:00 +0000

Project Overview

The Desi Maximalism LoRA is a specialized text-to-image style transfer model engineered to recreate the rich, dense, and vibrant aesthetic of mid-century Indian commercial art. It acts as a cultural bridge for generative AI, accurately reproducing the nostalgic visual language of 1940s–1985s South Asian matchbox labels, product packaging, magazine ads, and hand-painted film posters.

Sample Generations

Bournvita retro tin packaging

1 / 8

Model Specifications

Feature	Specification
Base Model	Qwen/Qwen-Image-2512
Fine-Tuning Method	LoRA (Low-Rank Adaptation)
Task	Text-to-image • Style transfer
Trigger Word	`desi-max`
License	Apache 2.0

Dataset & Curation Strategy

The model is trained on a highly focused dataset of 78 meticulously handpicked images of vintage South Asian commercial print. Each image was manually selected to represent a distinct visual sub-pattern. This strict curation keeps the dataset tight and entirely avoids the “style collapse” common in broader, noisier generative models.