Malhar Inamdar

⛰️ Among the hills of Mussoorie, Uttarakhand

Hey, I’m Malhar, a final year undergraduate at PICT, Pune, with a keen interest in machine learning research.

My interests include multilingual NLP, language modelling, interpretability and applications of machine learning in healthcare and climate.

Over the past year, I’ve worked as a research intern at Vizuara AI Labs on the project Regional-TinyStories, where we extended Microsoft’s TinyStories (2023) approach, which was originally for English, to Indian regional languages. We trained Small Language Models from scratch (5M–150M parameters) for Hindi, Marathi and Bangla, and developed a research framework for multilingual SLM evaluation, tokenizer analysis, translation data quality, and language complexity. Work accepted at IJCNLP-AACL Findings 2025. Earlier at Vizuara, I also worked with Mahindra Motors on a diffusion-based inpainting pipeline for automotive image editing.

At Froncort.AI, I built an RLHF pipeline that converted expert reviewer feedback into heuristic reward signals to improve LLM output quality under compute constraints, and architected a multi-agent system for regulatory document generation for medical devices.

I enjoy building projects that are meaningful and carry real-world impact. For the PICT Techfiesta hackathon 2025, I led the development of Vaidya Nidaan, an Alzheimer’s diagnostic platform combining CNN-based MRI classification, FSL biomarker analysis, Grad-CAM interpretability, and a multilingual RAG pipeline for medical report generation. The project was placed third among 400+ teams.

Feel free to reach out if you’re working in AI/ML or just want to discuss anything — I’d be glad to hear from you.

news

May 18, 2026	Joined Mastercard as a Summer Intern at the Pune Tech Hub. I'll be contributing to projects involving AI agents, automation, and developer tooling over the next couple of months.
Oct 25, 2025	My first Research paper Regional-TinyStories: A Small Language Model Framework for Evaluating Language Learning, Tokenizers, and Datasets — has now been accepted at *IJCNLP-AACL Findings 2025!*
Jul 24, 2025	Selected for the highly competitive Data Science: Probabilistic and Optimization Methods II program of International Centre for Theoretical Sciences - Tata Institute of Fundamental Research featuring lectures by leading experts on data science, probabilistic models, and optimization techniques.
Jun 14, 2025	Was selected to attend Microsoft Research India's Academic Summit 2025 virtually, a highly selective research summit held for researchers, academic professors and PhD, Master's and undergraduate students across India.
Apr 14, 2025	Regional Tiny Stories: Using Small Models to Compare Language Learning and Tokenizer Performance — preprint available on arXiv.
Feb 11, 2025	Our project focusing on the health of Alzheimer patients Vaidya Nidaan stood 3rd in Pune Institute of Computer Technology (PICT) Techfiesta Hackathon out of more than 400+ teams. Project Link
Apr 21, 2024	Stood 2nd (Runner-up) in Cretronix, a multi-stage contest involving electronic circuit design and microcontroller programming, conducted as part of Credenz 2024, the annual technical fest of PICT IEEE Student Branch.

publications

AACL

Regional-TinyStories: A Small Language Model Framework for Evaluating Language Learning, Tokenizers, and Datasets

Nirvan Patil*, Malhar Abhay Inamdar*, Agnivo Gosai*, and others

Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (IJCNLP-AACL 2025)

Abs URL PDF

Small, resource-efficient language models are pivotal for extending high-quality text generation to low-resource and regional languages—the true frontier of linguistic equity in AI. Yet research largely prioritises massive English-centric systems, leaving regional-centric (low-resource) language modelling underexplored, particularly how tokenizer design, dataset diversity, and linguistic structure shape the inference of Small Language Models (SLMs) under realistic computational and data constraints. We present Regional-TinyStories, a lightweight framework that treats SLMs as cost-effective stand-ins for LLMs, enabling rapid, variable-wise inference-based analysis. Extending TinyStories to Hindi, Marathi, and Bangla, we release datasets of 2M synthetic and translated stories per language and train over 20 SLMs spanning 5–157M parameters. Using this framework, we (i) uncover contrasts between form-oriented (grammar, fluency) and content-oriented (context, completeness, creativity) metrics; (ii) chart language-specific learning dynamics; (iii) rank tokenizers, showing Indic-specific Sarvam-1 outperforming SUTRA and generic Tiktoken (GPT-2) across all metrics; and (iv) demonstrate that dataset semantic quality (translation vs. synthetic) strongly governs downstream generation. Validation through an LLM-as-Judge ensemble (GPT-4o, LLaMA-3.3-70B) and a 100+ participant human study confirms these trends while exposing systematic score inflation in automated evaluations. Regional-TinyStories offers a reproducible path to benchmark tokenizers, datasets, and SLM designs for scalable, context-faithful generation in low-resource settings.