Projects
-
Mapwise
Introduced a map-based QA benchmark using Plotly, featuring 1,000 questions each for the USA, India, and China. Conducted evaluations of multiple LLMs, revealing a 20-50% performance gap compared to humans, and identified a tendency for hallucinations in counterfactual data scenarios.
Tech Stack: Python, LLM
2025 | | Code -
Partitioned Learned Bloom Filter
Implemented Learned Bloom Filter with GRU model, which partitions input to multiple Bloom Filters and results in 20x reduction in False Positive rate for same amount of memory as vanilla Bloom filter. Keeping False Negative rate to 0
Tech Stack: Pytorch, Python, Machine Learning
2024 | | Code | PDF -
Data Cleaning with LLM
Developed a system for large tabular data imputation from textual documents. Created QA pairs based on functional dependencies from tables and stored embedded documents in a vector DB. Fine-tuned T5-large using the QA pairs along extracted context using Retrieval Augmented Generation (RAG) methodology, achieving a 20% improvement.
Tech Stack: Python, Transformer, Vector DB, LLM, RAG
2024 | | Code -
Recommendation using Neuro Symbolic approches
Implemented three Neuro-Symbolic Machine Learning approaches for product recommendation based on user purchase history. First, the Apriori algorithm was utilized to identify frequently co-purchased items, which were then used to augment the training data. Next, aggregated contrastive loss was employed to differentiate between a user’s positive and negative reviews. Finally, an architectural adjustment was introduced by leveraging a hierarchical prediction method using the masking concept.
Tech Stack: Pytorch, Python, Machine Learning, Transformers
2024 | | Code | PDF -
Community Posting App
Developed a "Community Posting App" server using Ktor and an H2 in-memory database via the Exposed library. Supported routes for retrieving, creating, and deleting posts by ID, viewing posts since a timestamp, and fetching all posts. Used a phased approach with initial HashMap-based routing, later integrating persistent storage. Manually tested HTTP requests to ensure functionality.
Tech Stack: Kotlin, Ktor, H2
2024 | | Code -
Marble Rolling App
Created a "Marble Rolling App" using the gravity sensor to move a ball within screen bounds via offset in BoxWithConstraints. Used MutableLiveData or LaunchedEffect for dynamic updates, resolving axis and aliasing issues. Tested smooth motion with emulator controls at SENSOR_DELAY_GAME frequency.
Tech Stack: Kotlin
2024 | | Code -
Near By Shop Finder
An Rest-API which takes "pincode/postcode" as a input and gives a list of near by shops. An additional parameter is also provided which takes type of shop like resturants, grocery store and more. Error Handling, Unit testing and Swagger configuration are also perfomed. Google Maps Nearby API was used.
Tech Stack: Javascript, NodeJS, ExpressJS, Mocha, Chai, Swagger
2023 | | Code -
Visual Question Answering
Fine tuned BERT, ResNet, Vision-Language Transformer models, obtained Accuracy of 53.2% and F1-score of 21.0%. Converted VQA task to classification task (1000 classes), performed inference based on 7-Question and 3-Answer type.
Tech Stack: Pytorch, Python
2023 | | Code -
Neural Dependency Parser
Implemented a multi-class classifier and dependency parser for building transitional dependency tree of english sentences, obtained Labelled attachment score (LAS) of ≥ 75% for the Penn treeBank(PTB) dataset.
Tech Stack: Pytorch, Python
2023 | | Code -
Character Language Model
Developed a 2-layer LSTM based Language Model with ≈ 10M params, learned character level embedding from scratch. Applied Teacher Forcing during training, evaluation metric (Perplexity), Generated text without teacher forcing
Tech Stack: Pytorch, Python
2023 | | Code -
Text to Image Generation
Used two existing Generative Adversarial network (GAN) namely DCGAN and DFGAN, for generating realistic images from textual discription. We have used CUBDataset.
Tech Stack: Python, Pytorch, Numpy
2021 | | Code | PDF -
Sizzer - Cloth size predictor
Built an application to predict the size of upper body clothes from a captured image of a person using image processing techniques. Used an A4 size sheet as an reference object for calculating the real size of person upper body.
Tech Stack: Python, Opencv (cv2)
2021 | | Code -
SubGIN - Subgraph Isomorphism Detection using Graph Neural Networks
Designed a neural network architecture, in which, we proposed an attention-based graph pooling mechanism and an Interactive context attention layer. Applied CNN on derived feature matrices to learn node level similarity and Neural Tensor Network on graph embeddings to study their global relation. Evaluated the model on various graph datasets
Tech Stack: Python, Pytorch, networkx
2021 | | Abstract | PDF -
Sarcasm Detection
Implemented deep learning models such as Context based attention-network,LSTM and machine learning models such as support vectors machine, MLP on the n-grams and bert-vectorized features. We have used textual mustard dataset and have compared the importance of each input type such as context, character with actual utterance.
Tech Stack: Python, Pytorch, Keras, Scikit-learn, Pandas.
2020 | | Code -
Career Advisory System
Designed a career advisory system for the 4th year B.Tech. students at IIIT-Delhi using logic programming in Prolog. The system evaluate scores for 10 career domains based on the inputs provided by the user.
Tech Stack: SWI-prolog
2020 | | Code -
Traffic Sign Classification
Implemented deep learning architectures including MLP, CNN and RestNet50 for the classification task of 43 Traffic/Road signs. Achieved an Accuracy of 96.04% on test data, also implemented machine learning models including Logistic Regression, SVMs with different kernels (Linear, RBF, Polynomial with deg=3 ) and Random Forest on derived feature vectors.
Tech Stack: Python, Pytorch, Pandas, Scikit-learn.
2019 | | Code -
Engineer Chatbot
Developed a closed domain answering chatbot mobile app which provides details of project upon user query. The user could interact either by selecting from predefined options or by submitting a free text query. Each query is sent to a server where it is processed and fetches response from the database.
Tech Stack: Python, android-studio, NLTK, MySQL, socket programming
2019 | | Code -
Snake vs Block Game
Created an GUI game by following Object Orient Programming concept using JavaFX. The Project was selected among the top 10 projects for extra bonus marks from 120+ projects.
Tech Stack: Java, JavaFX
2018 | | Code