Research
Handwritten exam OCR
Vision-language model fine-tuned with GRPO on Gemma-4 to extract structured fields (roll number, page number, question number, content) from handwritten exam sheets.
View on Hugging FaceOUR WORK
From model to
meaningful application.
We build and adapt multimodal language models for Indian-context use cases, and make them efficient enough to deploy.
FLAGSHIP DIRECTION
India-focused multimodal LLM
In developmentOur core focus is building a multimodal (vision + language) LLM adapted for Indian-context use: handwritten and printed document understanding, Indian-language text, and practical deployment on accessible hardware. We build on open foundation models (Gemma-4 and Qwen3 MoE families) and focus on fine-tuning, reinforcement-learning-based adaptation, quantization and serving.
- Vision-language fine-tuning (LoRA, GRPO/GSPO-based training)
- Indian-context data and evaluation
- Efficient quantization for low-cost deployment
- Production serving with vLLM
Projects
Released
Qwen3.6-35B-A3B INT8
INT8-quantized Mixture-of-Experts model prepared for efficient vLLM serving.
View on Hugging FaceReleased
Gemma-4 quantization suite
FP8, W4A16 and NVFP4 quantized variants of Gemma-4 MoE models using llmcompressor.
View on Hugging FaceOngoing
Efficient serving with vLLM
Deployment recipes for quantized models on A100 and L40S GPUs, including cross-GPU accuracy comparison.
Our approach
Adapt
fine-tune open models for Indian-context tasks
Compress
quantize models to cut cost and memory
Deploy
serve with vLLM and test accuracy across hardware
THE NEXT CONVERSATION
Interested in collaborating or building with us?
Last updated: