I work at the intersection of AI systems, deep learning, and applied research. My work spans GenAI systems, retrieval-augmented pipelines, speech recognition, NLP fine-tuning, computer vision, generative modeling, and data-driven decision systems.
A Data Science and AI graduate from IIT Madras with a Minor in Generative AI and a Minor in Multimodal AI. My work is implementation-driven, spanning GenAI systems, RAG pipelines, speech AI, NLP fine-tuning, computer vision, reinforcement learning foundations, and applied data science.
I am especially interested in building reliable AI systems that can operate in real-world domains: healthcare, enterprise workflows, defense-oriented intelligence systems, and multimodal decision support. I approach each problem with an experiment-oriented mindset, drawn to emerging AI applications that demand both technical depth and engineering rigor.
I learn quickly by building: designing systems, running experiments, studying failure modes, and iterating toward reliability. My curiosity is particularly directed toward the convergence of language, vision, and speech in AI systems.
A selection of implementation-driven AI projects spanning generative systems, speech, NLP, and vision.
🏆 Best Software Engineering Project Award
GenAI System
AI-Powered Software Engineering System
AI Engineer · Team of 7
Developed and integrated a multi-module GenAI-powered academic and software engineering system designed to improve student learning, productivity, debugging, and academic workflow automation.
Students often spend significant time navigating lecture content, generating notes, revising weekly material, debugging code, and preparing for assessments. The goal was to build a unified AI-powered system that could assist across these workflows in a structured and usable way.
Approach
Worked as AI Engineer in a 7-member team, contributing to the design and integration of multiple GenAI modules. Built workflows around LLM APIs, prompt engineering, retrieval-based context grounding, backend orchestration, and modular AI pipelines.
Modules Built
AI chatbot with contextual assistance
Video / lecture summarisation
Week-level summarisation
Topic notes generation
Practice question generation
Mock quiz generation
Coding assistant for error explanation
Topic recommendation support
Technical Depth
The system involved transcript processing, content chunking, LLM prompting, vector database retrieval using ChromaDB, structured outputs, backend API integration, and response reliability handling. The focus was on building usable AI workflows that fit into an academic software platform, not just generating responses.
Key Learning
Learned how to move from isolated GenAI prompts to integrated AI systems involving data flow, context management, reliability, modularity, and user-facing workflows.
Speech AI
Automatic Speech Recognition Systems
ASR Pipelines · Self-supervised Models
Built and optimised end-to-end ASR pipelines using self-supervised speech representation models including wav2vec2.0, HuBERT, and Whisper, with focus on audio preprocessing, CTC decoding, and GPU-efficient training.
Speech recognition systems must handle variable-length audio, noisy samples, inconsistent sampling rates, and alignment between audio signals and text tokens. The goal was to build stable ASR pipelines capable of preprocessing speech data, training/using models, decoding outputs, and analysing transcription quality.
Approach
Worked on audio preprocessing, waveform normalisation, tokenisation, batching, padding, alignment strategies, and GPU-efficient training workflows. Implemented CTC-based decoding and encoder-based ASR architectures while experimenting with representation learning strategies for transcription robustness.
Technical Depth
Designed preprocessing pipelines for variable-length audio, sampling-rate normalisation, memory-conscious batching, tokenisation, dynamic padding, model loading, and decoding. Used Hugging Face Transformers and PyTorch-based pipelines for model experimentation.
Key Learning
Gained practical exposure to speech representation learning, sequence modeling, CTC alignment, encoder-based architectures, inference stability, and ASR system design.
NLP Fine-Tuning
Google Gemma Fine-Tuning with LoRA / PEFT
LLM Adaptation · Parameter-Efficient Methods
Fine-tuned Google Gemma large language models for domain-specific NLP tasks using parameter-efficient fine-tuning techniques including LoRA and PEFT workflows within the Hugging Face ecosystem.
PyTorchHuggingFacePEFTLoRAGemmaInstruction Tuning
Problem
Large language models often need adaptation for specific tasks, domains, or response styles. Full fine-tuning can be computationally expensive, so parameter-efficient techniques are useful for improving model behavior while reducing resource requirements.
Approach
Designed training pipelines involving dataset preprocessing, prompt formatting, tokenisation, batching, LoRA configuration, training configuration tuning, and inference testing. Focused on instruction-following behavior, domain adaptation, response consistency, and efficient deployment.
Technical Depth
Worked with Hugging Face Transformers, PEFT, PyTorch, tokenizer pipelines, prompt engineering, instruction-tuning workflows, and GPU-accelerated training environments.
Key Learning
Developed practical understanding of LLM fine-tuning, parameter-efficient adaptation, transformer behavior, prompt formatting, inference optimisation, and model robustness evaluation.
Computer Vision
4× Image Super Resolution
Computer Vision Competition · CNN Architectures
Developed deep learning based image super-resolution systems focused on reconstructing high-quality images from low-resolution visual inputs using convolutional neural network architectures and residual learning.
PyTorchCNNResidual LearningPSNR/SSIMPerceptual Loss
Problem
Image super-resolution requires recovering fine-grained spatial detail and improving perceptual quality from low-resolution inputs. The challenge is to improve sharpness and texture consistency without introducing artifacts.
Approach
Built 4× image super-resolution pipelines using CNN-based architectures, residual learning concepts, feature extraction, perceptual loss optimisation, and adversarial training concepts.
Technical Depth
Worked on image resizing, normalisation, patch-based training, augmentation, GPU-efficient workflows, training stability, and inference optimisation. Evaluated results using PSNR, SSIM, and perceptual quality analysis.
Key Learning
Gained exposure to image restoration, CNN architectures, perceptual learning, adversarial optimisation concepts, and computer vision experimentation.
Generative AI · Vision
GAN-Style Architecture for Generative Image Modeling
Generative AI Competition · Adversarial Training
Developed and trained GAN-style architectures for generative image modeling tasks, focused on producing realistic image outputs from encoded datasets, with careful attention to training stability and mode diversity.
Generative image modeling requires stable adversarial training and maintaining diversity while improving realism. GANs are difficult to train due to mode collapse, generator-discriminator imbalance, and unstable loss behavior.
Approach
Built end-to-end generative pipelines involving dataset preprocessing, image decoding workflows, training data organisation, augmentation, generator-discriminator optimisation, and GPU-accelerated training.
Technical Depth
Worked on encoded image shard datasets, efficient loading, normalisation, batching, latent-space behavior, generator capacity tuning, discriminator balancing, regularisation strategies, and FID-style evaluation concepts.
Key Learning
Gained practical exposure to adversarial learning, latent representation learning, training stabilisation, GPU-intensive experimentation, and generative modeling workflows.
Data Science
Business Data Management · Native Chefs
Applied Data Analysis · B2C Business Insights
Analysed real-world business data from Native Chefs, a B2C home-cooked food delivery business, to identify operational and revenue-related insights around unpaid orders, dish performance, and customer behavior.
PythonPandasMatplotlibEDABusiness Analytics
Problem
The business needed better visibility into unpaid orders, dish-level performance, customer ordering behavior, and revenue leakage.
Approach
Performed data cleaning, preprocessing, descriptive statistics, exploratory analysis, pivot-based summaries, dashboarding, and business interpretation.
Focus Areas
Revenue leakage through unpaid orders
Dish-level performance analysis
Customer ordering behavior patterns
Revenue trend identification
Operational improvement recommendations
Key Learning
Gained practical experience in turning raw business data into decision-support insights, bridging the gap between data exploration and actionable business recommendation.
Advanced Coursework
Academic Training
Areas of Interest
Research Focus
I am especially drawn to AI systems that combine research depth with real-world application potential: healthcare diagnosis support, enterprise intelligence workflows, multimodal reasoning systems, speech interfaces, and high-stakes decision-support systems.
⬡
AI Systems
End-to-end intelligent pipelines and modular AI architecture
⬡
Healthcare AI
Diagnosis support, clinical decision systems, medical NLP
⬡
Multimodal AI
Vision-language integration and cross-modal reasoning
Sequential decision-making and policy optimization
⬡
Computer Vision
Image understanding, restoration, and generative vision
⬡
Defense-Oriented AI
AI for high-stakes decision systems and strategic intelligence
Recognition & Experience
Highlights
Award
Best Software Engineering Project
Recognised for the AI-Powered Software Engineering System developed as part of the IIT Madras curriculum, a multi-module GenAI platform serving academic workflows.
Project Role
AI Engineer
Led AI module design and integration across a 7-member engineering team, responsible for GenAI pipelines, RAG architecture, prompt engineering, and backend orchestration.
Academic Credential
Minor in Generative AI · IIT Madras
Completed a focused specialisation in Generative AI covering LLMs, deep learning practice, and the mathematical foundations of generative systems.
Academic Credential
Minor in Multimodal AI · IIT Madras
Completed coursework spanning Large Language Models, Speech Technology, and Deep Learning for Computer Vision — covering the three pillars of multimodal AI.
Additional Learning
Digital Marketing & AI in Business
Explored performance-driven digital marketing and AI applications in business growth, building awareness of how AI intersects with commercial decision-making and user acquisition.
Let's Connect
Get In Touch
Open to research collaborations, AI engineering discussions, and opportunities in applied AI, deep learning systems, and intelligent product development. Feel free to reach out.