Ankit Kumar

Aevrilo

Sat, 17 Oct 2026 00:00:00 +0000

Overview

What is Aevrilo?

Aevrilo is a research workspace that helps people understand how materials behave when stretched. It takes measurements from a tensile-testing machine and brings the curves, specimen information, analysis, and research notes into one place.

A researcher can inspect a test, record the specimen’s dimensions, compare samples, and produce a report while keeping the connection to the original measurements.

Why I built it

A material-property number means little without knowing how it was obtained. Which sample did it come from? Were the dimensions measured or assumed? Which part of the curve was used? Did someone review that particular analysis?

Language models from scratch

Sun, 15 Sep 2024 00:00:00 +0000

Overview

Understanding what happens inside a language model

Language models turn a sequence of words into a prediction about what comes next. This project explores that process by building the main components directly, rather than treating a transformer as a black box.

I worked from attention and embeddings up to a GPT-style model, then explored two practical adaptations: following instructions and classifying SMS messages as spam or ordinary messages.

The project connects the mathematics of language modelling with the software needed to prepare text, train a model, save its state, and generate output. It is an educational implementation of established architectures.

Reinforcement learning experiments

Mon, 19 Sep 2022 00:00:00 +0000

Overview

Teaching an agent through consequences

Reinforcement learning is about learning what to do by observing the consequences of an action. This project explores that idea through simple control tasks and Tank Dodge, a custom game in which an agent must keep moving to avoid approaching enemies.

The agent receives a reward for surviving and a penalty for collision. Over training, it learns a movement policy from those signals rather than from a list of correct moves.

Modelsmith

Wed, 01 Jun 2022 00:00:00 +0000

Overview

Bringing machine learning into a usable application

Modelsmith brings different machine-learning architectures into a shared web interface. Instead of moving between separate notebooks, users can explore image classification, image generation, prediction, and language-processing demonstrations in one application.

I developed it with Ashwani Ahlawat. The work combined model implementations with a Django application and technical explanations of the ideas behind them.

Technical detail

One application, several model families

Separate Django apps organise CNN classification, GAN generation, neural-network prediction, and NLP workflows. The project covers more than ten architectures, including custom ResNets and generative adversarial networks.

Omarchy Project Manager

Mon, 12 Oct 2026 00:00:00 +0000

Overview

Getting back to work faster

Opening a project usually means more than finding its folder. I also need an editor, a terminal, the current Git state, and the right task for that particular codebase.

Omarchy Project Manager puts those actions in a keyboard-driven desktop overlay. Fuzzy search helps find a project, and frequently or recently used projects move closer to hand.

It is a tool built around my own Linux workflow: reducing the small interruptions involved in switching between different kinds of software projects.

Tensile Analysis Platform

Thu, 19 Mar 2026 00:00:00 +0000

Overview

Making laboratory measurements easier to understand

A tensile-testing machine records how a material responds as it is stretched. Turning those readings into useful comparisons involves more than plotting a curve: spreadsheets need to be connected to specimens, relevant properties calculated, and differences between groups interpreted.

I built this Streamlit dashboard for a University of Stuttgart research laboratory to bring those steps into one workflow. It combines raw-data handling, stress–strain analysis, material-property calculations, and comparison tools.

Helpdesk ticket API

Fri, 26 Sep 2025 00:00:00 +0000

Overview

Keeping support requests from getting lost

A helpdesk needs more than a form for reporting a problem. People need to find their own requests, support agents need an assigned queue, and unresolved work needs a way to reach the right people.

This project is a ticket-management backend built around that workflow. It organises requests by status, priority, and assignee, with different responsibilities for requesters, agents, and administrators.

The aim is to connect everyday support operations with the backend decisions that make them possible: access control, searchable records, background work, and a clear API.

Nuclei image segmentation

Thu, 22 Jun 2023 00:00:00 +0000

Overview

Identifying individual structures in microscopy images

This project segments cell nuclei in microscopy images. Instead of assigning one label to an entire image, it predicts which pixels belong to the structures of interest.

A Django interface connects image input to segmentation output, bringing the model into an application that can present its result visually. The work joins model development with the engineering needed to make inference accessible.

Technical detail

A custom PyTorch U-Net uses an encoder–decoder structure and double-convolution blocks. The training pipeline uses the 2018 Data Science Bowl nuclei dataset and includes preparation, checkpointing, pixel-accuracy checks, and intersection-over-union utilities.

AI Agent Planner

Sun, 21 Sep 2025 00:00:00 +0000

Overview

Turning an intention into a starting plan

AI Agent Planner accepts a goal written in everyday language and produces a suggested plan. It adds web-search and weather context where relevant, and keeps a history so earlier plans can be revisited.

The idea is to reduce the blank-page problem between knowing what someone wants to do and deciding how to begin. The application proposes steps; the person remains responsible for choosing and carrying them out.

Reading recommendation system

Thu, 20 Feb 2025 00:00:00 +0000

Overview

Learning what a reader might enjoy next

This experiment explores reading recommendations using patterns in user interactions. Rather than comparing content only by keywords, it learns relationships between readers and the items they engage with.

The goal is to turn interaction history into a preference signal that could help order reading suggestions. The current project is a model experiment, rather than a deployed recommendation product.

Technical detail

Preprocessing maps users and items to numerical indices and prepares ratings for a custom PyTorch Dataset. The model learns separate user and item embeddings, concatenates them, and passes them through a multilayer network to predict an interaction score.

Video detection & crowd alerts

Wed, 04 Jun 2025 00:00:00 +0000

Overview

Making video activity easier to interpret

This project explores how object detection can turn video into summaries and event alerts. It identifies what appears in sampled frames and looks for a crowd that persists over time, rather than reacting to one isolated detection.

The work also compares lightweight detection models, exploring the practical balance between inference speed and the information returned.

Technical detail

YOLO and OpenCV support frame sampling and object detection. The crowd-event rule requires at least three people across five consecutive analysed frames, introducing a temporal condition before an alert is raised.

GitHub Webhook Dashboard

Wed, 09 Jul 2025 00:00:00 +0000

Overview

Seeing repository activity in one place

This dashboard turns GitHub activity into a readable event feed. It brings pushes, pull requests, and merges together so a viewer can follow what changed, who acted, and which branches were involved.

It is an example of event-driven integration: an action in one service becomes a stored event and then a visible update in another application.

Technical detail

GitHub sends webhook payloads to a Flask receiver. Relevant fields are stored in MongoDB, and a JavaScript dashboard polls for new activity every fifteen seconds.

Court case search dashboard

Fri, 01 Aug 2025 00:00:00 +0000

Overview

Finding a case and returning to it later

The court-case dashboard offers a single place to search Delhi High Court cases and inspect available order or judgment links. Users provide a case type, number, and year, then revisit earlier successful queries through a history dashboard.

Its purpose is to connect a structured search with a readable result and a useful record of past lookups.

Technical detail

A Flask application handles the search and history pages. Selenium and BeautifulSoup retrieve case information from the source website, while structured results are stored for later viewing.

Docker-backed Python editor

Wed, 28 May 2025 00:00:00 +0000

Overview

From a browser editor to running code

This prototype lets a user write Python in a browser, run it, and see the output without moving between an editor and terminal.

The project explores the connection between an interactive coding interface and a separate execution environment—the basic workflow behind a browser-based coding tool.

Technical detail

Flask serves the editor and accepts execution requests. The Docker SDK passes submitted code to a container-based workflow, then returns the result to the browser.

Video subtitle extraction & search

Thu, 13 Oct 2022 00:00:00 +0000

Overview

Making video content searchable through subtitles

This application connects video uploads with background subtitle extraction and search. Its purpose is to make spoken content easier to find without asking users to manually watch through every uploaded video.

The longer-running processing work is separated from the web interface, allowing upload and search to remain distinct parts of the experience.

Technical detail

Django handles the application pages, while Celery coordinates asynchronous extraction through Redis. The processing task invokes a subtitle extractor and reads the resulting subtitle data with Pysrt.

Fake-news detection hackathon

Tue, 11 Oct 2022 00:00:00 +0000

Overview

Putting text classification in front of users

This first-place UpGrad hackathon project connects a news-text classifier to a Django web application. Users can submit text through an interface instead of interacting directly with a model or notebook.

The project explores how a language-processing model becomes part of a usable application. Its prediction is a classification signal, not an independent determination that a claim is true or false.

Technical detail

TensorFlow handles text preprocessing, sequence padding, and recurrent-model prediction. The classification module also includes retraining functions.

Hyprland Session Saver

Wed, 29 Oct 2025 00:00:00 +0000

Overview

Returning to a familiar workspace

Reopening applications after a desktop restart is only part of recovering a working session. The terminal’s directory, an editor session, and the media someone was watching also matter.

Hyprland Session Saver records applications and their workspaces, then restores them with application-specific context. It grew from the practical needs of my Linux desktop workflow.

Technical detail

Python session handling maps applications back to their workspaces. Alacritty support recovers a working directory and detects a Neovim session; MPV handling records the media file and restores playback paused.

Vinland healthcare prototype

Sat, 03 Sep 2022 00:00:00 +0000

Overview

Bringing several health-model demonstrations into one interface

Vinland is a collaborative hackathon prototype that brings thirteen health-related machine-learning demonstrations into an authenticated web application. It explores how different prediction and image-classification tasks can share an accessible interface.

I worked with Krish Yadav, Gaurav Patel, SwakshWar Ghosh, and Prakriti Verma. The application is a prototype for exploring model integration, not a clinically validated diagnostic service.

Technical detail

Django apps organise the interface and authentication, while separate model modules handle individual inference workflows. The project combines PyTorch and TensorFlow models across tabular prediction and image-classification tasks.

Machine learning project collection

Tue, 26 Apr 2022 00:00:00 +0000

Overview

Exploring different ways to apply machine learning

This collection brings together my earlier learning and internship prototypes. It spans animal-intrusion alerts, driver-drowsiness detection, handwriting recognition, plant-disease classification, text similarity, and question tagging.

The projects vary from notebook experiments to small applications. Together, they show how I explored different problem types and began connecting models to interfaces and services.

Technical detail

The collection covers computer vision, NLP, prediction, and application integration. Streamlit interfaces, voice utilities, MongoDB-backed tools, and S3 storage workflows appear alongside training experiments.

Multilingual RAG System

Mon, 01 Jul 2024 00:00:00 +0000

Overview

Asking questions across documents and languages

Large PDFs and books can be difficult to navigate when someone needs a specific answer. That becomes harder when the question and the source material use different languages.

This project combines document search with language-model generation. It retrieves relevant passages first, then uses that context to support an answer rather than relying only on the model’s existing knowledge.

Technical detail

The pipeline combines PDF text extraction, OCR for image content, text splitting, Hugging Face embeddings, and FAISS retrieval through LangChain. Sample material spans Bengali, English, Urdu, and Chinese.

AquaQuest

Wed, 01 Nov 2023 00:00:00 +0000

Overview

Learning about water through play

AquaQuest is an educational platformer for children aged five to six. It introduces the water cycle and groundwater conservation through an interactive adventure.

Rather than asking children to learn only from explanations, the game places ideas inside playable levels, then reinforces them through short questions and prompts.

Technical detail

The game uses Godot and contains four interactive levels. Concept scenes, mini-quizzes, and tooltips connect progression with the learning material.

Attention Is All You Need, Explained from the Ground Up

Tue, 20 Oct 2026 10:00:00 +0530

Imagine translating “The animal did not cross the street because it was tired.” Before translating “it,” you need to connect it to “the animal.” Its meaning depends on something several words away.

Now imagine doing that for every part of a sentence: deciding which earlier or later pieces help interpret each token, then using those relationships to write a translation.

That is the problem at the heart of Attention Is All You Need, the 2017 paper by Ashish Vaswani and colleagues that introduced the Transformer. Its central proposal was to build a sequence-to-sequence model around attention, removing recurrence and convolution from the main architecture. Attention existed before this paper; the architecture built around it was the important step. Read the original paper.

This guide starts with an everyday explanation, then works through the actual operations. You do not need to know Transformers already. Familiarity with lists of numbers and basic Python will help; the sections on shapes make the matrix operations explicit.

By the end, you should be able to trace one token through the model, explain the attention equation, distinguish the three attention patterns, and understand what a training step is trying to learn.

Tensile Strength Analysis

Thu, 19 Mar 2026 10:00:00 +0530

Introduction

Tensile testing is one of the most fundamental tests in materials science and mechanical engineering. It measures how a material responds to being pulled apart until it breaks. This analysis transforms raw machine data—force and displacement—into meaningful mechanical properties that engineers use for design and quality control.

This guide explains the scientific methodology behind the Tensile Test Analysis Dashboard, covering everything from basic stress-strain calculations to advanced techniques like sliding window regression for E-Modulus determination.

Understanding Large Language Models: Mathematics and Architecture

Sat, 24 Jan 2026 10:00:00 +0530

Introduction

Large Language Models (LLMs) have revolutionized artificial intelligence and natural language processing. From GPT to BERT, these models demonstrate remarkable capabilities in understanding and generating human language. But what makes them work? In this post, we’ll explore the mathematical foundations and architectural principles that power these sophisticated systems.

The Foundation: Neural Networks

At their core, LLMs are built upon artificial neural networks. A neural network consists of layers of interconnected neurons, each performing a simple mathematical operation:

What is Cloud Computing?

Mon, 03 Feb 2025 19:53:33 +0530

Introduction

Cloud computing is the delivery of computing services—including servers, storage, databases, networking, software, and more—over the Internet (“the cloud”). Instead of owning your own computing infrastructure or data centers, you can rent access to these services from providers like Amazon Web Services (AWS), Microsoft Azure, or Google Cloud Platform.

Using a public cloud is about using other people’s servers to run your digital workloads. In a sense, there’s no significant difference between running a software application on servers hosted in your own office versus locating it within Amazon’s infrastructure. In both cases, you need to make sure you’ve got sufficient compute, memory, network, and storage resources. In both cases, fast deployments and avoiding over-provisioning are key goals.

Image Gallery

Sat, 25 Jun 2022 18:35:46 +0530