I’m exploring offline reinforcement learning for data-center cooling-system optimization: learning operational policies from previously collected data rather than relying exclusively on online interaction. My interests in this project include sequential decision-making, data-driven policy learning, reward design, evaluating learned policies, and the practical constraints of deployment. I’m drawn to the gap between reinforcement-learning theory and reliable real-world applications.
With mini-modern-llm, I’m exploring how to build a small language model and understand its components rather than treating language models as black-box APIs. I’ve trained a tokenizer and saved it as part of the project. The topics I’m exploring include tokenization, text representation, transformer architectures, language-model training, and implementation. The project is at https://github.com/hey-shiv.
AgentAlign-Lab is an ongoing project in which I’m exploring agent alignment and related AI-system concepts. I’m developing it incrementally, aiming to understand the underlying ideas and build the implementation systematically.
Polymind is a project and development workspace that I’ve worked on and reorganized as part of my broader exploration of AI and software engineering.
I’m Shivashant, a fourth-year Computer Science and Artificial Intelligence & Machine Learning (CSE-AIML) undergraduate at VIT-AP University. I’m an aspiring ML Engineer interested in building intelligent systems, understanding modern AI architectures, and turning research ideas into practical products.
- My interests span machine learning and deep learning; LLMs, transformers, inference, and AI agents; reinforcement learning and offline RL; RAG and memory systems; AI engineering and model deployment; audio processing and music information retrieval; generative AI; open-source projects; research-driven product development; and developer tools and AI-assisted workflows.
My technical interests include the mathematical foundations and assumptions behind machine learning, feature representations, optimization, and evaluation, as well as how classical ML connects to modern deep learning. I have a strong foundation in neural networks and am continuing to explore architectures, representation learning, training dynamics, attention, transformers, generative models, and deep learning for audio. I want to understand LLMs beyond prompt engineering and API integration: I’m interested in tokenization, training, inference, KV caching, attention, RAG, context management, short- and long-term memory, agents, tool calling, evaluation, and inference optimization, especially the engineering decisions behind production-grade applications. I’m also interested in reinforcement learning for sequential decision-making, including Markov Decision Processes, policy optimization, value functions, exploration and exploitation, offline RL, reward design, and evaluation. My broad interest in music and audio AI includes Music Information Retrieval, signal processing, harmonic and rhythmic representations, audio embeddings, similarity and cover-song identification, music generation, neural audio models, music transformers, and multimodal understanding; it is not limited to any particular instrument. This connects to my work on music audio and ai work.
I prefer learning from first principles rather than blindly following tutorials or copying implementations. My usual sequence is to understand the problem, study its mathematical and theoretical foundations, read research papers and technical explanations, examine existing implementations, build an implementation myself, test and evaluate it, and then identify its limitations and possible improvements. I prefer theory before implementation, but I want theory connected to practical engineering; structured learning paths should avoid unnecessary repetition so I can spend more time on advanced concepts. I value research papers, technical books, engineering blogs, official documentation, university course materials, well-written technical articles, and lectures from established researchers and institutions, including resources from MIT and other leading universities and engineering teams. I care about why an architecture or algorithm works, why it was selected, its trade-offs, and where it might fail. For each project, I want to understand the problem, the appropriateness of the approach, alternatives and assumptions, how results can be validated, limitations, and what production use would require. I enjoy comparing existing implementations with research findings. A project is not successful merely because it runs or has an attractive interface: technical correctness, reproducibility, testing, and validation matter. I use AI coding assistants as engineering tools, not replacements for my technical understanding.
I see modern AI engineering as broader than training models: inference, retrieval, memory, context management, evaluation, deployment, and optimization are all important to useful systems. I want to understand the infrastructure around models and the shift from standalone language models toward systems combining retrieval, tools, memory, and autonomous workflows. I believe useful AI products pair capable models with carefully designed systems and user experiences, and that AI should support creative work without sacrificing the quality of its outcomes. I want to be a product-oriented engineer, exploring real, useful products rather than stopping at prototypes or academic demonstrations. My product-development approach is to identify a meaningful problem, research existing solutions, understand target users, design the technical architecture, build core functionality, develop the user experience, test and improve the product, and explore sustainable business models. I’m particularly interested in AI-native products where intelligence is a meaningful part of the experience, and in how coding agents can accelerate product development and reduce repetitive engineering work.
I have explored an automated short-form content workflow that would turn a topic and script into a finished video while maintaining a deliberate, consistent visual identity. I want production to become repeatable through automation, but my approach is to establish creative direction and quality standards first and automate gradually. I have considered whether this workflow could become a paid product, potentially with a credit-based system, an initial free experience, and paid plans; it remains an idea under exploration. My creative-technology interests include automated video production, AI-assisted editing, motion design, visual storytelling, typography, design systems, creative automation, AI-powered content tools, and productizing creative workflows. I care about visual quality and consistency and dislike generic interfaces that resemble other AI-generated websites. I want design decisions to be intentional across typography, color, spacing, layout, motion, information hierarchy, interaction, and visual identity, and I prefer a recognizable design language over unexplained changes in style. Technical products should also have a thoughtful user experience.
Python is my primary language for machine learning and AI development; I use Java for programming and data structures and algorithms, and C++ for programming and problem-solving. My environment and tools include macOS, VS Code, Git and GitHub, Conda and Python virtual environments, Homebrew, Jupyter notebooks, and Kaggle. I use GitHub to maintain and develop projects, and have worked with Git, repository organization, testing, and documentation. I’m interested in reading source code, contributing to open-source software, building on existing research, and exploring developer tools that improve engineering productivity on macOS. I value projects that are technically meaningful, independently examinable, and reproducible. I use Hugging Face, Claude Code, Google Gemini, Gemini Deep Research, research papers, and technical documentation; I’m working toward practical experience with PyTorch. I have also explored web development to make technical projects accessible, including Vite, static web applications, GitHub Pages, Vercel, and frontend integration for ML projects. My GitHub is https://github.com/hey-shiv, my portfolio is https://hey-shiv.github.io/, and my X account is https://x.com/hey_shiv; my GitHub username is hey-shiv and my online handle is hey_shiv.
My immediate goal is to become an ML Engineer with strong practical and theoretical foundations, gaining industry experience while continuing to develop research capabilities, contribute to meaningful open-source projects, and demonstrate actual engineering ability through my work. My areas of interest include ML engineering, deep learning, LLM engineering, AI infrastructure, reinforcement learning, audio and music AI, and applied AI research. Long term, I want to work at the intersection of AI, music, and engineering: understand the science behind intelligent audio systems, contribute to their engineering, and eventually build products in this space. I’m interested in organizations such as Google DeepMind, including its music research efforts, and Suno. I want to combine deep technical expertise, applied AI research, production-grade engineering, creative technology, product development, and music and audio intelligence. My research-to-production ambition is to read a paper, understand its mathematical foundations, implement its central ideas, optimize the system, and turn it into something people can use. Current areas I’m exploring include inference from first principles, modern LLM engineering, RAG and memory architectures, agents and agentic workflows, model optimization and deployment, reinforcement learning, music information retrieval, audio representation learning, generative music, AI-assisted creative production, and automated software products.
I approach AI from both engineering and research perspectives: I enjoy understanding how systems work internally, reading technical papers, experimenting with implementations, and building systems that solve meaningful problems. I’m especially interested in the engineering challenges that appear when research concepts meet real-world constraints.
I want to become a strong ML Engineer working on modern AI systems, rather than limiting myself to traditional machine-learning applications. I choose projects to demonstrate genuine technical understanding and to have the potential to become useful products; I’m not interested in building projects just to fill a portfolio.
I’m particularly interested in the intersection of AI and music: the engineering behind intelligent music systems, generative audio, and next-generation creative AI tools. I also see opportunities for AI-powered products in audio and visual storytelling. This creative-technology interest connects with my work in music audio and ai work.
- I want to understand AI systems across their lifecycle: mathematical foundations, architectures, training and fine-tuning, inference and optimization, retrieval and memory, agentic workflows, evaluation and reliability, and production deployment.