Projects

Note: The list here includes all my open-source projects that I worked on. I have not included the projects that I worked on as part of my job.

Asymmetric Dense Passage Retrieval

Technologies:
python
simpletransformers
pytorch
docker

Dense passage retrieval is the algorithm behind search query engines like Google. Given a query, find all the relevant documents that have matching vector. Asymmetric Dense passage retrieval, uses 2 BERT-like models, one that is light weight, trained for building search vectors, and another bigger model that is iteratively trained on any new document added to the database. I could improve the DPR accuracy by 4% across all major benchmarks.

Project =>

Super Knowledge Graph Engine

Technologies:
python
pytorch
pykeen, numpy for data handling
docker

A knowledge graph model that can genaralize to unseen relationships without any training. Makes use of RMPI graph network along with BERT to learn embeddings for relationships and attempts to generalize from just training on the KG once. I wrote my M.sc thesis on Knowledge graph relation prediction, and this was the outcome of various experiments I did.

Project =>

Automatic Blog Writer

Technologies:
python
FastAPI
c++
docker

Generate blog like articles, given a seed keyword and instructions on tone, writing style. This is currently in maintenance with no active development. It includes multi threaded (pipeline) architecture, where each pipeline generates one part of the blog. Uses Large Language Model like GPT-2.5, GPT-3 and also fine-tuned StableDiffusion model to generate images. I still use some of the pipelines actively for my other projects.

Project =>

Heading Grabber Chrome Extension(Under development)

Technologies:
Typescript
Javascript
Node

Simple chrome extension to grab headings and format it to export easily.

Project =>

Counterfactual Generative Networks

Technologies:
Python
docker
pytorch
pytorch-lightning
wandb

A [Re]Science FAIR project, where we implemented the paper "Counterfactual Generative Networks" by Yoon et al. The paper proposes a method to generate counterfactuals for a given input. We implemented the paper and also extended it to generate counterfactuals for a given input and a target class. We also implemented a method to generate counterfactuals for a given input and a target class, while also ensuring that the counterfactuals are semantically similar to the input. Our analysis got accepted to the [Re]Science FAIR 2021 conference.

Project =>

ScoopStats

Technologies:
Svelte
Typescript
Vite
Node
.NET (Az functions, auth and storage)

My portfolio is living here, but this is hosted on Azure with a back-end containing a headless CMS-like system delivering the blog articles. Backed by cosmos DB and Azure blobs for storing the `.md` files.

Project =>