portrait of Vishal Mandrai

"I believe successful AI is not defined by model size or complexity, but by its ability to solve the right problem efficiently, economically, and delightfully for the end user."

Hi, I'm Vishal.

A Data Scientist and AI Developer passionate about building intelligent products that solve real-world problems. With expertise in Statistics, ML, Deep Learning, and Gen-AI, I transform ideas into scalable, production-ready solutions.

My work spans the entire AI lifecycle—from research and data engineering to model development, deployment, and monitoring. I build AI-powered Web Applications, recommendation systems, computer vision pipelines, data platforms, and LLM-driven products using modern engineering practices, cloud infrastructure, and MLOps principles.

Above all, I prioritize user experience and measurable impact, building technology that people genuinely find useful and businesses can confidently scale.

Download My Resume
My Top Skills
Recent Projects
BookHead AI App

BookHead AI App

BookHead AI is an AI-powered web application designed for both readers and librarians. Instead of manually searching through book lists or cataloging shelves one book at a time, you simply upload a photo of books arranged on a bookshelf and let AI handle the rest.

Built using open-source AI models and tools.

FastAPIQdrant DBCeleryRedisFaster R-CNNPaddleOCRGemma-2B
Gurugram Real Estate Explorer App

Gurugram Real Estate Explorer App

An Interactive Data Science App designed for Homebuyers, Investors, Developers, and Policy makers to explore the real estate dynamics of Gurugram with ease. Built on the latest data from leading housing platforms, this app serves both as a market snapshot and a tool for in-depth analysis.

PythonStreamlitPlotlyFoliumXGBoostOptunaSelenium
Personal Website

Personal Website

Built and customized a personal portfolio website using an existing Node.js and Tailwind CSS template, leveraging GitHub Copilot to add new features and overcome unfamiliar technologies through AI-assisted development.

ReactNext.jsTailwind CSSGitHub Copilot
VoxForm AI

VoxForm AI

Building a comprehensive AI based application for filling forms on voice inputs. For organizations, to quickly build forms using an interactive pick and drop interface. Fill forms at scale with pace. Voice models and GPT like LLMs work under the hood for error corrections and adjustments while form filling. Interactive dashboard to visualize collected form data and download reports.

Application coming soon.

FastAPIMySQL DBCeleryRedisGPT-5Whisper
All Projects →
Recent Posts

Inference Is Not Execution: Why Some Problems Defeat Every LLM

At some point, no matter how capable the model is, the task becomes impossible — not because the model lacks knowledge, but because it simply cannot perform enough computation before producing its next token. This realization changed how I think about Large Language Models. Their biggest limitation may not be hallucination or missing knowledge. It may be something much more fundamental: every LLM has a fixed computational budget during inference.

Regularization in ML | Elastic-Net | Part 3/3

Last post of Regularization trilogy. Here we'll dive into "Elastic-Net Regularization". Often called the best of both worlds. Creates flexible, robust and interpretable solution. Jump in to see how combination of L1 and L2 penalty in loss function, work togther to make possible a great model.

Geometric Intuition of how sign(𝑤ⱼ) aligns with sign(z) in Lasso Regression

While deriving the optimal solution for Lasso Regression via Coordinate Descent based optimization. Their comes a defining point where solution turns to Soft Thresholding Operation for calculating optimal 𝑤ⱼ, because sign(𝑤ⱼ) aligns with sign(z). This post helps with geometric intuition of how sign(𝑤ⱼ) aligns with sign(z) that enables Soft Thresholding Operation for calculating optimal 𝑤ⱼ.

Regularization in ML | Lasso | Part 2/3

I hope you liked Part 1 of this trilogy. Next we'll dive into "Lasso Regularization". The ruthless feature selection algorithm and a great simplifier. Prevent overfitting. Automatic feature selection by vanishing the coefficients of less important features. L1-penalty is the main reason for shrinkage. But how? It cannot be solved via Gradient Descent Algorithm needs a special optimization algorithm. Jump in to see which one?

Regularization in ML | Ridge | Part 1/3

A technique in ML that prevents overfitting in complex ML system, by adding a penalty to the Loss Function. Helps improve a model’s generalization ability. Introduced “Penalty” discourages the model from learning overly complex patterns in data, and instead learn the most visible pattern, hence later perform well on unseen data.