2025

Nebula (preview)

Build Local Apps on the fly right in your web browser. Github Nebula is an open source agentic browser built with Electron & Chromium from scratch.

Snapgenie

My first SaaS - to turn product snaps to professional photographs

Qwen Model Family Integration in Keras OSS

As a consultant for the Keras team, I led the integration of the state-of-the-art Qwen model family into KerasHub. This work made multiple versions of this powerful generative model, including its advanced Mixture-of-Experts variant, accessible to the global Keras community.

• Qwen 3 & Qwen 2.5: For both the Qwen 3 and Qwen 2.5 models, my contributions included implementing the core backbone, custom attention mechanisms, and building a robust checkpoint conversion pipeline.

• Qwen-1.5 Mixture-of-Experts (MoE): The integration of the Qwen-1.5-MoE model presented unique challenges, particularly its high memory footprint. To address this, I engineered a memory-efficient MoE layer, adapting techniques from other advanced implementations to ensure the model was performant and accessible on standard hardware.

Mixtral OSS

Through this work I developed support for Mixtral model in Keras Hub, making it accessible to community.

2024

JMLR Paper on Keras Multi-framework Models Publication

I co-authored a foundational paper for the Journal of Machine Learning Research (JMLR) detailing the multi-framework architecture of KerasCV and KerasNLP. The work showcases how these libraries were engineered to seamlessly support TensorFlow, PyTorch, and JAX.

• Title: KerasCV and KerasNLP: Multi-framework Models
• Published: Journal of Machine Learning Research (Volume 25, 2024)
• My Role: Co-author, contributing to the framework development and experimental validation.

Youtube video YouTube

Data hack summit Talk

I gave this talk at Data Hack Summit by Analytics Vidhya. It covered CUDA C fundamentals and core CUDA concepts like threads, blocks, and memory hierarchies. The session guided attendees through writing their first CUDA kernels, aiming to build the skills needed to optimize deep learning workloads.

2023

I contributed to open source (mainly Keras) heavily throught the year. Gave a lot of talks travelling across the country.

Google Summer of Code at TensorFlow OSS

My Google Summer of Code project was centered on expanding the KerasNLP library with foundational generative models. This experience evolved from implementing a single model to contributing core components, porting functionalities to Keras 3, and adding a series of highly-demanded models to the ecosystem.

• Pioneering Foundational Models:
  • GPT-NeoX: My main project was the end-to-end implementation of GPT-NeoX, including its preprocessor. Foreseeing the industry’s shift, I proactively ported the model to Keras 3 to ensure multi-backend support for TensorFlow, PyTorch, and JAX.
  • Llama 2: Responding to immediate community excitement, I applied my learnings from GPT-NeoX to quickly contribute the Llama 2 model to KerasNLP shortly after its release.
• Developing Core API Components:
  • Generic Rotary Embedding: I identified that the Rotary Embedding layer for GPT-NeoX was a fundamental component for most modern LLMs. I took the initiative to refactor and ship it as a generic, reusable Keras layer, strengthening the core Keras API.
  • New Keras Ops: To support the models, I implemented and contributed the rsqrt op and added Root Mean Square scaling to the LayerNormalization layer, both of which are essential for modern generative architectures.
• Enhancing Existing Models & Frameworks:
  • Albert & XLM-RoBERTa: I expanded the capabilities of existing models by contributing missing task layers, including the Classifier and Masked LM heads for Albert, and the Masked LM head for XLM-RoBERTa.
  • Beam Sampler Port: To deepen my understanding of decoding strategies, I ported the Beam Sampler to Keras 3, contributing to the framework’s multi-backend compatibility.
  • Testing Improvements: I significantly improved the CI/CD pipeline by speeding up RoBERTa testing by ~3x and adding new TPU tests for backbones.

Talks Talk

In later half of 2023, after I finished my GSoC, I decided to give back the knowledge I earned from OSS, to community. So I flew 11 cities throughout India giving talks and workshops on various topics related to NLP. Few of them are listed below:

1. KerasNLP: From Words to Wisdom, on October 7, 2023 at DevFest New Delhi’23 .

2. Modular NLP Workflows with KerasNLP, on September 29, 2023 at Google Developer Groups, Seattle.

3. Supercharging Keras with WandB, on September 23, 2023 at TensorFlow User Group Mumbai.

4. GSoC Success Secrets: Cracking the Code to OSS Excellence, September 10, 2023 at National Institute of Technology, Warangal.

5. Rethinking LLM Design with KerasNLP, on August 26, 2023 at TensorFlow User Group Hyderabad.

6. KerasNLP for Starters, on August 20, 2023 at TensorFlow User Group Durg

7. Taking KerasNLP on GenAI Ride on July, 23, 2023 at TensorFlow User Group, Kolkata

8. Re-imagining Keras in the evolving ML ecosystem on July 16, 2023 at Google I/O Extended’23 New Delhi.

Julia Contributions OSS

Metalhead is FluxML’s (a julia deep learning package) official model zoo. I contributed the UNet image segmentation model to Metalhead.jl, Julia’s core computer vision library. By implementing the architecture with reference to the official PyTorch version, I made this popular model for semantic segmentation natively available to the FluxML community, fulfilling a key feature request for the library.

Misc Bug Fixes OSS

• Fix RotaryEmbedding import
• Fix Autograph error with perplexity metric
• Fix ModuleNotFoundError keras_nlp.models.xlnet
• Add compute_output_shape to tokenizer
• Default compilation for Albert, Distilbert, Roberta MaskedLM
• Call super.config() in BartBackbone’s get_config()
• Add API exports for tokenizers documented on keras.io
• Add API exports for metrics documented on keras.io
• Add API exports for samplers documented on keras.io
• Add API exports for models documented on keras.io
• Move from_preset to base tokenizer classes
• Add an add_prefix_space Arg in BytePairTokenizer

2022

This year I mainly focused on Kaggling. It also marked my venture into OSS world.

CNN From Tensorflow to Pytorch OSS

In 2022, Deepchem Project decided to migrate from tensorflow to pytorch. I contributed to this mission by porting their generic, scalable CNN models and layers to torch. I also fixed couple (1, 2) of bugs in the library that have gone unnoticed.

Neural Ordinary Differential Equations OSS

[Deepchem - Neural ODE tutorial 📚]

This tutorial demonstrates how to implement Neural Ordinary Differential Equations (Neural ODEs) within the DeepChem library by leveraging the torchdiffeq package. It explains how, instead of using discrete layers, a neural network can parameterize the derivative of a system’s hidden state.

The guide walks through embedding a torchdiffeq ODE solver inside a PyTorch neural network to model continuous-time dynamics, making it possible to predict the future states of dynamic systems within the DeepChem ecosystem.

G2Net Detecting Continuous Gravitational Waves Kaggle

I scored a bronze medal in this competition. The goal of this competition was to find continuous gravitational-wave signals. I developed a model sensitive enough to detect weak yet long-lasting signals emitted by rapidly-spinning neutron stars within noisy data. Here is one of the competition notebooks that shows my approach to problem.

ICR - Identifying Age-Related Conditions Kaggle

The goal of this competition was to predict if a person has any of any medical conditions. Here are two notebooks that I submitted for the competition. One of them used an ensemble based technique and the other used just LigthGBM. I used optuna library for hyperparameter optimization.

Support safe serialization while pushing to Huggingface Hub from H2O llmstudioOSS

PR

Deepmind - optax OSS

This is how I started open source. My first-ever contribution was a jax library called optax by google deepmind.