..

About me

About Me

“The best way to understand something is to build it, break it, and then build it better.”

Hey there, I’m Aayush!

Welcome to my digital space! I’m a machine learning researcher with a passion for making machines speak. This is where I share my thoughts, experiments, discoveries, and everything that doesn’t fit into a formal research paper.

Think of this as my public notebook - a mix of technical deep-dives, project walkthroughs, personal notes, and honest reflections on the world of ML.

What I’m Up To Now

Currently working as an ML Researcher at Sarvam, where I’m focused on:

  • Audio Foundation Models: Building the next generation of voice-to-voice systems
  • Music Generation: Developing AI systems that can compose and create music
  • Research: Pushing the boundaries of what’s possible in audio AI

Previously, I was a Founding Data Scientist at Nurix where I built Dialogue Manager (Semantic-VAD) for Audio Bots.

My Journey Through ML

Professional Experience

Sarvam | ML Researcher

Currently building audio foundation models for voice-2-voice applications and developing music generation systems. My work focuses on pushing the boundaries of what’s possible in audio AI architectures.

Nurix | Founding Data Scientist

Built a SOTA Dialogue Manager that detects turn-ends, pauses, and interruptions.

ConvoZen.AI | Associate Data Scientist

Developed AI-driven contact center intelligence solutions including an Indic ASR model with, co-developed the ConvoZen LLM from dataset creation to fine-tuning, and delivered an Indic NER model.

Research Internships

Elucidata | NLP Intern

Built a Federated Weak Learning framework for unlabeled data with Multi-NER models achieving 88% accuracy. Delivered a Dockerized system demo that was accepted for AMIA ‘22 in Washington, DC.

University of Utah | NLP Research Intern

Built custom RoBERTa + BiLSTM models that tap external knowledge graphs (ConceptNet, WordNet). Improved Infotabs NLI benchmark by 7.5% and published the Trans-KBLSTM paper.

Envestnet | Yodlee India | Data Science Project Trainee

Instituted prior framework using BiLSTMs, Bayesian networks, ARIMA and statistical methods for wealth recommendations. Built end-to-end data agnostic system with self-evaluation criteria that enhanced problem-specific model choice, yielding 15% increase in recommendations consumed.

Projects & Research

Trans-KBLSTM | 2022

External knowledge-enhanced transformer for tabular reasoning

  • Best Paper Award at DeeLIO-ACL, Dublin
  • Enhanced NLI performance using ConceptNet and WordNet

Structural Transitions of Dehydrin Protein | 2025

Investigated structural changes in DHN1 (dehydrin) protein from Zea mays using spectroscopic techniques (CD, fluorescence, FRET). The study revealed disorder-to-order transitions in response to temperature and chemical stressors, providing insights into plant stress tolerance mechanisms. Published in Biochemistry with contributions to protein engineering and stress response analysis.

Academic Background

B.Tech in Biotechnology | IIT Guwahati | 2019-2023

  • Convener, MLRW’22
  • Member, IITG.ai & Consulting and Analytics Club

Interesting journey: Started with biotechnology, ended up teaching machines to understand and generate audio.

Why This Website Exists

This space serves as my digital notebook where I:

  • Document experiments and learnings
  • Share thoughts about AI and audio research
  • Write tutorials and technical guides
  • Keep personal notes that might be useful to others
  • Explore creative applications of ML

Whether you’re here for the technical content, looking for inspiration, or just curious about my journey - welcome!

Let’s Connect

Thanks for visiting my corner of the internet! I hope you find something interesting in my notes and thoughts about the fascinating world of machine learning.