Building
Projects
“My Experiments with ML and stuff.”
Synapse: Configuration Language for Memory Systems
A domain-specific language (.mnm files) and runtime for building memory systems for AI agents. Define schemas, event handlers, queries, and update policies declaratively, which compiles into a multi-backend runtime.
Domain
Stack
Rust, SQLite, Qdrant, Neo4j
DSRs: A rewrite of DSPy with Rust in mind
DSRs is a Rust rewrite of DSPy that preserves its high-level abstractions while giving us native async and zero-cost safety. Built a rust crate that can work with multiple LLM providers and use a block-like mechanism for building LLM pipelines.
Domain
Stack
Rust, Tokio, Async
DSPy: Programming—not prompting—Language Models [Core Maintainer]
Built support for native Data Loading [DataLoaders], Synthetic Data Generation [Synthesizer], Agent Optimization [Avatar]. Adding multithreading and multiprocessing support to allow batched requests in DSPy. Building dev friendly docs.
Domain
Stack
Python, PyTorch, Multiprocessing
Scaling Hyperbolic Embeddings for Dendocentric Learning
Working on training, ablations, and quantization of Hyperbolic BiEncoders to enable fast retrieval through nano-dendrites. Created complete infrastructure for end-to-end training and evaluation of hyperbolic retrievers through projection methods and Reimann optimization.
Domain
Stack
PyTorch, Hyperbolic Geometry
Pirate: Distributed Training and Evaluation Framework for Retrievers
Built a library to enable efficient and up-to-date triple mining techniques, training and evaluation framework to remove any bottleneck arising from library usage in IR research. Implemented various methods for triple mining pretraining and training of retriever models.
Domain
Stack
Python, PyTorch, Distributed Training
RingSAGE: Improving Graph-Level Tasks with Virtual Nodes for Molecular Rings
A novel Graph Machine Learning algorithm that introduces virtual nodes to explicitly represent molecular rings during the message-passing stage of Graph Neural Networks (GNNs). This approach addresses the limitation of traditional GNNs in uniformly treating molecular graphs, thereby missing unique ring properties crucial for determining molecular characteristics like hydrophobicity, solubility, and bioavailability. The model outperformed standard GNNs on the ZINC dataset for molecular drug solubility prediction.
Domain
Stack
PyTorch, Graph Neural Networks, Cycle Detection
Sift.ai - Personalized and Automated Resume Screening Tool
Sift AI is a powerful, easy-to-use tool designed to simplify the hiring process for businesses. When companies have too many job applications to look at, Sift AI steps in to help. It quickly goes through all the resumes and picks out the best candidates for the job. This means HR teams and company founders can save a lot of time because they only need to focus on interviewing the most promising applicants.
Domain
Stack
LangChain, HuggingFace, ChromaDB, ReactJS, TailwindCSS, AWS, Node.js
Terse.ai - AudioBook Summarization and Sharing Platform
Terse.ai is a new audio book summarization and sharing platform. Terse allows users to quickly summarize audiobooks in a few minutes, creating a text-based summary that can be easily shared with others. You can read and hear summaries made public by users around the globe. This makes it easy for people to get a quick overview of the most important points from an audiobook, without having to listen to the entire thing.
Domain
Stack
HuggingFace, TailwindCSS, PyTorch, Pytube, Flask, AWS, Moviepy, HTML/CSS/JS
WeCare - Torn down Medicine Strip Prediction
The WeCare was created to help elderly people find and identify the medicine strip from torn strips. Wecare is designed with a simple and easy-to-use interface that allows users to quickly scan the torn strips and match them with the correct medicine. Once the strip is identified we also provide the target disease of the medicine along with side-effects in case the user forgot.
Domain
Stack
OpenCV, Flask, Bootstrap, NLTK, Tesseract OCR, Requests, BeautifulSoup, HTML/CSS/JS
Knots - Cyclone Speed Prediction from Satellite Images
Knots is a web app that uses satellite images to predict the intensity of a cyclone. It provides information on cyclone speed, it's severity, and countermeasures so users can make informed decisions about their safety. The app can be used by weather forecasters to provide early warning of a cyclone's speed and intensity. Along with that it provides a weather card that gives you real time details of your area including wind speed, temperature, humidity etc.
Domain
Stack
PyTorch, Flask, Pillow, WeatherAPI, Bootstrap, GCP, HTML/CSS/JS
Social Distance Surveillance with Physical Alerts and Streaming
During the early stages of COVID-19, I built a security system that can monitor CCTV footages and tell which frame has violation of social distancing, it also provides live video streaming and physical alerts via alarm and NodeMCU. This system was created when social distancing was newly implemented. However, I wanted a system that would be able to be of practical use and not just a POC. Hence, we included an alarm system and streaming option so it could function like and an actual security system.
Domain
Stack
PyTorch, Detectron2, Flask, NodeMCU, Firebase, Scipy, Numpy, OpenCV, HTML/CSS/JS, Bootstrap
SeqFetch-My Unwatched Anime Sequel List
Keeping track of Anime was hard for me and I wanted to make sure I've completed all the sequels of animes I've watched. So I developed SeqFetch a web app that takes your MyAnimeList ID and provides a list of sequels you missed based on the anime you have completed! SeqFetch is a simple but powerful app that helps you keep track of your favorite series and ensure you have a good time watching them.
Domain
Stack
Requests, BeautifulSoup, Pandas, SQLite, HTML/CSS/JS, Bootstrap
ChattY - Chat UI/Backend as a Service
What if we tell you that you can add a chat interface to your product without any hassle? Forget about us giving you the design, we'll provide you with everything! Literally everything! A backend and a UI, just connect your user database to ChattY and let the messages fly. That's what Chatty does it's an npm package that gives your app a chat UI kit to design or use a premade Chat UI and a backend handle to handle chats.
Domain
Stack
Next.js(Frontend), React.js(UI-Kit), Node.js, WebSockets, Axios, MongoDB, TailwindCSS, TypeScript, JavaScript
MSRNet-LowLight Image Enhancement
Implemented this paper. MSRNet is an network proposed by team at Huazhong University of Science and Technology, Wuhan that directly learns an end-to-end mapping between dark and bright images. The architecture was chosen keeping in mind the architecture should be lightweight with decent quality.
Domain
Stack
PyTorch
Attention is All You Need
Implemented this paper. Attention is not a concept this proposes, how a standalone implementation of attention for Seq2Seq task is what This paper proposes which this introduces. This became the foundation for many new age models for wide range of tasks not limited to NLP.
Domain
Stack
PyTorch
AutoCal- Voice Driven Handwriting Machine
AutoCal is a voice-driven handwriting machine that listens to you and turns it into text written on paper. With AutoCal, we give you an Applet you can you to speak into, then transmit the recognised text to AutoCal and let it write! What makes this special is that it was made on an extremely low budget. We used discarded DVD writers that served as the base of the movement system. We used NodeMCU instead of RaspberryPi to keep cost to minimum.
Domain
Stack
Blynk, NodeMCU, Motors
Smart City Demo
A mega project consisting of many mini-projects depicting how a modern day Smart City will look like. We implemented the following:-
Domain
Stack
Blynk, NodeMCU, Arduino, MQ-2(Smoke), DHT-11(Temperature), HC-SR04(Ultrasonic), Servo Motors, HC-SR501(Proximity)
PiLock-Face Recognition Door Unlock via RaspberryPi
PiLock is a face recognition door unlock prototype deployed onm RaspberryPi to identify individuals and unlock the door. The system is designed for use in homes, offices, and other locations where access control is desired. The PiLock can be configured to only unlock for individuals added.
Domain
Stack
PiCamera, RaspberryPi, PyThon, OpenCV, LBPH, Motors
Deep Contrast Enhancement for OLED Displays
Implemented this paper. In this method, the power consumption is constrained by simply reducing the brightnessa certain ratio, whereas the perceived visual quality is preserved as much as possible by enhancing the contrast of the image using a convolutional neural network (CNN).
Domain
Stack
PyTorch
Salient Object Detection API
A FastAPI based API that enables developers to detect and identify salient objects in images and use it remove the background from them. API makes it easy to add background removal capabilities to your applications without having to worry about the complexities of building and training your own machine learning models.
Domain
Stack
FastAPI, PyTorch, TorchVision
Line Following Robot
A simple autonomously guided robot that follows a line drawn on the ground to either detect a dark line on a white surface or a white line on a dark. This Project was created as a learning exercise for Arduino which became a project we participated in many competitions with.
Domain
Stack
Arduino, LDR, Motors, Motor Control Board
Auto Garbage Disposal System
An automated garbage disposal system that helps make taking out the trash easier and more efficient. The system includes a wifi-controlled trash car prototype to automatically move to garbage bins of each house to collect garbage, and an automated garbage disposal bin to automatically dispose of trash when the trash truck arrives.
Domain
Stack
NodeMCU, IR Sensor, Motors, Motor Control Board, Servo Motors
Hindi Image Captioning via VIT+GPT2
This is an encoder-decoder image captioning model made with VIT encoder and GPT2-Hindi as a decoder. This is a first attempt at using ViT + GPT2-Hindi for image captioning task. We used the Flickr8k Hindi Dataset available on kaggle to train the model. This model was trained using HuggingFace course community week, organized by Huggingface.
Domain
Stack
HuggingFace, PyTorch, HFTrainer
Weather Changing in Image using CycleGANS
Implemented a model that changes the weather in an image using CycleGANS. With it, you can easily change the weather in any image to match your given image conditions. Whether you want to create a winter wonderland or a summer paradise, it makes it easy to change the weather in any image.
Domain
Stack
PyTorch
Automated Parking System
A system that automatically finds the license plate number from the vehicle footage and register it in the csv file. This saves time and manpower for the security department, by doing the tedious task of finding and inputting license plate numbers into the csv file for keeping parking records. This can be used on tolls as well were the work is of similar nature but the amount of vehicles is high, making it slow to do it manually.
Domain
Stack
SQLite, Pandas, Tesseract-OCR, OpenCV
LiveStats-You Statistics PlayGround
Learning is one thing and seeing what you learned in action is another thing. Keeping that in mind I created this package where you can input data and validate few statistical theorems like bessel's correction etc. Just input the data into any module you want to play around with and see the math come to life.
Domain
Stack
Numpy, Matplotlib, Seaborn
MouseBit-Direct DataLoader Creation on Custom Data
DataLoader creation itself is a one liner task however the process can comprise of many boilerplates to solve this issue I came up with MouseBit to make the whole process itself a one liner. Just input the data and config, it'll tackle everything from transformation to DataLoader creation on it's own.
Domain
Stack
PyTorch
OMR Scanning Application
A web app that takes input a filled OMR sheet, answer key and student details and outputs the score of the student by evaluating the OMR. Click the picture of OMR sheet and done. No need to get a full fledged OMR machine to do this task, no need to sit and check. Just practice and evaluate!
Domain
Stack
Flask, OpenCV, Numpy
Projects for Learning
I'm heavily driven by project based learning, applying what you learned exposes you to problem you would've never faced. I've created projects for everything I learned be that image classification on Oxford Flower Dataset, text classification on twitter dataset, neural networks from scratch, text summarization, Face generation with GANS, etc.
Domain
Freelance Projects
I was fortunate enough to have gotten a wide range of domains to work on in my freelancing. Many of these projects were domains I never explored but they became my bridge to it. I've worked a lot with people in higher education and research, helped them out with paper implementation and improvements to the existing work. I've helped people deploy their models onto their environment and help them come up with optimizing latency as well. Aside from that I've build a many of projects from scratch for people as well.
Domain