cv

Graduate Engineer Trainee

Havells India Limited

Working at Havells India Limited as a Graduate Engineer Trainee.

• Developed and deployed multilingual ASR models with a focus on Indian-accented English, optimized for resource-constrained microcontroller ESP32-S3, meeting stringent memory and compute requirements.
• Designed a novel 5M-parameter ASR model achieving 0.3 Word Error Rate (WER) and 0.15 Character Error Rate (CER) on curated 650 hours of Indian English datasets including IndicTTS, NPTEL English, Svarah, while maintaining a compact 6MB model size suitable for real-time edge inference.
• Trained and benchmarked multiple architectures, including Conformer, Conmer, and stateless RNNT models for performance comparison and deployment suitability.
• Contributed to the TensorFlow Lite Micro repository by implementing custom operators essential for on-device inference, enabling successful deployment of ASR models on ESP32-S3.
• Created frame-level phoneme-aligned datasets using Azure Speech Services, supporting accurate acoustic modelling. Designed and integrated robust data augmentation pipelines, including colored noise, impulse response convolution, random gain, and background mixing to enhance model generalisation.

IoT Intern

Havells India Limited

Worked on IoT projects at Havells India Limited.

• Developed a compact 450KB end-to-end speech-to-intent model for slot filling and intent classification task, achieving 93% intent accuracy and 94% slot accuracy on proprietary dataset.
• Reproduced and implemented the approach from the paper “A Low-Latency ASR-Free End-to-End Spoken Language Understanding System”, enabling real-time spoken command interpretation in edge environments.

CircuitBot

CircuitBot is an offline-capable, edge-deployed voice assistant designed for controlling IR-based home appliances without internet connectivity. The system integrates ASR and a local LLaMA 3.2-1B-Instruct LLM to interpret and execute user voice commands in real-time.

• IoT
• TinyML

Techdocs

A Code documentation tool, powered by SOTA LLMs like Llama2 and WizardLM-70b which analyzes the code with up to 90% accuracy and easily documents Python code bases inplace with the help of ast module.

• NLP
• Machine Learning

Bachelor of Technology

Vellore Institute of Technology, Bhopal

Computer Science and Engineering

• Operating Systems
• Computer Networks
• Data Structures
• Database Management Systems
• Machine Learning
• Artificial Intelligence

Top 10 finish in Meta LLama Hackathon

Reskill and Meta

Achieved Top 10 finish in Meta LLaMA Hackathon, selected from 150+ selected teams.

Honorable Mention

Docker

Honorable Mention in Docker AIML Hackathon competing against 2300+ participants.

Machine Learning Lead

GDSC VIT Bhopal

Mentored 300+ students by delivering tech talks and organizing hackathons.

Volunteer

Unnart Bharat Abhiyan

Digitized data intake for 4 adopted villages.