Malhar Inamdar

Hey, I’m Malhar, a third-year undergrad at PICT, Pune, with a profound interest in machine learning research and developing systems that can help solve complex problems and have large scale impact.

My research interests center on impactful applications of machine learning such as across climate, healthcare, language modelling and reasoning capabilities. I’m particularly drawn to building systems that improve accessibility in low-resource settings and interpretability that improves model trust and reliability - whether that means developing diagnostic tools for underserved communities or creating language models, frameworks for mutlilingual settings.

What I’ve worked on

Vaidya Nidaan — Led the development of an Alzheimer’s diagnostic platform that integrates CNN-based analysis with medical imaging using FSL biomarker identification (hippocampal volume, white/gray matter ratios). The system employs GradCAM for visual interpretability and includes a multilingual RAG pipeline that generates structured medical reports grounded in research literature. This work secured third place among 400+ teams at PICT Techfiesta 2025.

Medical Reasoning with Phi-3 — Fine-tuned Phi-3-mini-4k-instruct on the MedQA-USMLE dataset to teach structured clinical reasoning. Using QLoRA, I trained the model to generate explainable diagnostic chains-of-thought while engaging only 0.44% of its parameters. To explore preference-based alignment further, I implemented a GRPO trainer from scratch in PyTorch, working through the practical challenges of low-precision optimization.

Regional TinyStories — At Vizuara AI Labs, I conducted research on multilingual language modeling under Dr. Raj Dandekar, extending Microsoft’s TinyStories (2023) work for Indian regional languages. I trained Small Language Models (2M–150M parameters) from scratch for Hindi, Marathi, and Bengali. We developed a novel framework for the development and analysis of SLMs, tokenizer performance, linguistic complexity, machine translation performance and demonstrated that a 54M parameter model could achieve GPT-3.5-comparable results on short-story generation - a promising direction for accessible, high-quality models in Indian languages. Work accepted at AACL Findings 2025 (IJCNLP).

At Froncort.AI, I implemented a RLHF pipeline that converted expert reviewer feedback into heuristic reward signals, improving LLM output quality by 35% while working within tight compute constraints compliant with international regulatory and ontological standards; and architected multi-agentic system for regulatory document generation workflows.

When I’m Not Training Models..

I’m probably reading books, scrolling through interesting ml papers, listening music at odd hours or wandering through the hills. I’m drawn to the philosophical questions that emerge from AI’s capabilities, but I’m most energized by the opportunity to build systems that genuinely help people. Music often accompanies this process — it’s become a quiet companion that helps me think through problems and find clarity.

If you’re working on something in healthcare, climate AI or language modeling or just ml in general, feel free to reach out! I’d be glad to hear from you.