Atlas — Multilingual Foundation Models for Indian Languages
A 7B-parameter open model family trained on 22 Indian languages. Built for downstream fine-tuning across legal, medical and educational corpora.
01 · Impact
Six years of disciplined inquiry, measured not by output alone but by the systems, scholars and ideas that endured.
02 · The Lab
SPARKS Lab is a research ecosystem at the intersection of machine learning, cognitive systems, and the human disciplines that give them meaning.
Mission
To advance the science of artificial intelligence through rigorous research, open knowledge, and systems that serve real human needs at scale.
Vision
A future where intelligence is interpretable, equitable, and instrumental in solving the most consequential problems of our time.
03 · Ecosystem
We don't research in silos. Each domain feeds the others — perception informs language, language informs reasoning, reasoning informs the systems we deploy into the world.
Explore all research →04 · Featured work
A 7B-parameter open model family trained on 22 Indian languages. Built for downstream fine-tuning across legal, medical and educational corpora.
Transformer-based fusion of EHR, neural imaging and temporal vitals achieving 94% precision on validation cohorts across three partner hospitals.
A pruning and distillation framework cutting transformer inference latency by 40% on ARM and RISC-V edge hardware with no accuracy compromise.
05 · Timeline
Established at the Department of CSE, NIT Tiruchirappalli with three founding researchers.
Lab's foundational paper on attention-driven knowledge distillation accepted at NeurIPS.
Major government funding awarded for explainable AI in clinical decision support.
Sparse-inference framework granted patent by the Indian Patent Office.
Founding member of the National Mission on AI translational research consortium.
Released 7B multilingual foundation model under permissive open license.
06 · Scholarship
Abdullah A., Sharma R., Krishnan V., et al.
We show that structured knowledge graphs can constrain LLM hallucinations through dynamic node retrieval and iterative self-correction.
Iyer M., Abdullah A., Patel S.
A decentralized framework for adaptive vision-model updates across constrained IoT hardware using sparse graph structures.
07 · Network
Joint research, shared infrastructure and cross-continental scholar exchange programs with 24+ partner organizations.
Join SPARKS
Join a community of researchers, innovators and engineers creating real-world impact through AI — at one of India's premier institutes of national importance.