CV

B.Tech

• Relevant Courses:
  • Gained a comprehensive understanding of key areas through courses like Data Structures, Algorithms, Operating Systems, Big Data, and Software Engineering.
• Skills Acquired:
  • Acquired hands-on experience with multiple programming languages, such as Python, and C++, along with proficiency in Data Analysis and Machine Learning.
• Projects and Research:
  • Actively involved in projects focused on Machine Learning with Natural Language Processing and Computer Vision.
  • Developing a comprehensive benchmark for performance-improving code-generation analysis of LLMs.
  • Building a system to help users try different clothes virtually by just uploading the image.

Research Fellow

• Developing LLM with a focus on code generation, instruction-following, and reasoning, working on post-pretraining, online and offline RL methods.
• Led large-scale model training on multi-GPU clusters, optimizing distributed training workflows and fine-tuning to achieve high-performance code synthesis and execution.

Open Source Engineer (Numpy)

• Implementing an extended precision floating point Data Type
• Extending current level of hardware support for extended precision data types. It no longer looks like x86 is the only game in town.
• Packaging and distribution of a multi-platform python package that includes C extensions.

Data Science Intern

• Assisted in the development and deployment of predictive models to minimize infrastructural discrepancies and mitigate potential revenue losses. This proactive approach resulted in a 25% uptick in user engagement.
• Developed automation pipelines to auto-pre-process user engagement data and curate monthly reports and insights on personalized recommendations.
• Utilized statistical analysis and machine learning techniques to identify infrastructure issues and developed mitigation methodologies.
• Made core contributions to Scaler modules in Data Science and Machine Learning.

Open Source Research Engineer

• Contributed significantly to research projects enhancing code generation and large language models for code, and evaluated mathematical reasoning strategies.
• Collaborated on StarCoder, a 15.5B parameter multi-lingual code-gen model trained on 1 trillion tokens, surpassing OpenAI code-Cushman-001 model.
• Collaboratively developed OctoPack, leveraging Git commits for instruction tuning and leading to the creation of CommitPack data, Extended HumanEval benchmarks and Octo-duo models.

Machine Learning Engineer Intern

• Developed Deep Learning models for business market growth from domains of Computer Vision and Language modelling.
• Engineered a comprehensive API to automate the conversion of a pretrained model into ONNX format and its seamless deployment via Nvidia Triton Server, culminating in a 12% reduction in VM load.
• Developed custom CUDA kernels to accelerate 3D image processing for medical scans. Achieved a 2x speedup in segmentation tasks compared to existing CPU-based solutions, thereby improving the overall throughput of the system.
• Worked on state-of-the-art techniques like DreamBooth, Dichotomous Image Segmentation.
• Supporting model deployment team in model code analysis and optimizations for DeployX

Applied Machine Learning Instructor

• Mentored final year undergraduate students in their course of Machine Learning.
• Evaluated and improved the projects developed by students.

Virtual Clothing Assistant(275+ ⭐️)

• Allow user to try on different clothes virtually, without going to any trial room.
• Used ResNet101 and UNet to segment the cloth and model respectively and pose estimation using openpose then final predictions are done using the PyTorch implementation of VITON.
• SSIM (Structural Similarity Index Metric) of 0.895

MIRA - Multimodal Image Reconstruction with Attention *

• MIRA is a multimodal transformer for Text/Image to 3D reconstruction just using single 2D image of object within seconds
• It uses pre-trained DINO-V2 as encoder and custom triplane decoder that learns to project features on triplane via cross-attention and model the relations among the spatially-structured triplane tokens via self-attention, camera features are modulated within the decoder.
• Trained by minimizing the difference between the rendered images and ground truth images at novel views, without the need for excessive 3D-aware regularization or delicate hyper-parameter tuning.

S.E.A.R.C.H (Systematic Engine for Analyzed Retrieval and Contextual Handling)

• S.E.A.R.C.H is a production-ready intelligent open-source agent that can process factual information, and documents across web, providing interactive, personalized assistance.
• Backed by Zephyr-7B-beta model with efficient post-quantized fine-tuned on self-rag dataset using LoRA, to improve cross-domain performance.
• Production-ready, scalable and efficient Retriever-Augmented Generation system, Retrieving process backed with Keyword extraction + ReRanking.
• Custom subsequent-query engine allows generating multiple in-context related queries to intelligently cover all possible scenraios.
• In-built code-interpreter allows efficiently solving reasoning and critical tasks much better than ChatGPT3.5

3D Cervical Spine Segmentation and Multi-Vertebrae Fracture Detection

• Developed "3D Cervical Spine Segmentation and Multi-Vertebrae Fracture Detection" to automate cervical spine fracture detection from CT scans, aiming to match radiologist accuracy and ensure timely medical interventions.
• Leveraged the RSNA 2022's Kaggle Competition dataset; implemented a two-stage model, with the first stage focusing on 3D segmentation using EfficientNet + UNet to create binary masks for the cervical vertebrae (C1-C7).
• The second stage combined Convnext + LSTM for classification. Extracted 15 even slices per vertebrae sample by the z-dimension and added the predicted segmentation mask as an additional channel to distinguish between multiple vertebrae.
• Created a 3D reconstruction of spine to visualize the bone and corresponding fracture vertebrae.
• Achieved a Multilabel Dice Score of 0.92 for segmentation and used a Modified BCE loss with a score of 0.365 for classification.

PythonCoder [PyTorch]

• A custom 60M params GPT-like casual code generation model trained and evaluated on large Python codebase for 10k steps.
• Self-implemented GPT2 architecture using MultiQueryAttention allows fast large batch inference and FlashAttention allows 3x faster calculation of attention values with context window of 1024 tokens
• Training process leverages the entire context length by concatenating the sequences into a one big chunk, enabling faster training and better consumption of compute resources

Image Style Transfer using CNN [PyTorch, Pillow, OpenCV]

• It can impose the texture/style of one image on the other without disturbing the content of the other image.
• Used pre-trained VGG-19 network to extract the features from multiple convolutional layers for style and content then merged them while maintaining the content and style tradeoff.

Reinforced Edwin Lander [PyTorch, Stablebaselines3, Gym]

• It's a custom Gym environment built to train RL models on the Edwin Lander game to land via following the rules
• Current model is trained on PPO algorithm and attained the perfect landing score by training up to 1M steps.

Multilingual Name Entity Recognition [PyTorch, Huggingface]

• XLM-RoBERTa fine-tuned in German, French, and Italian languages enables robust Zero-Shot Classification.
• It achieves the expected f1-score of 0.8638 on validation set and gives decent f1-scores on low-resource languages.
• Thoroughly processed and analyzed using a custom Error Analysis pipeline.
• Ensures reliable classification across multiple languages through meticulous analysis and error handling.

Reasoning With StarCoder

• User-driven playground for interacting with StarCoder, assessing its reasoning capabilities.
• Evaluation strategies including PAL, TA, and MathPrompter.
• Model generates code based on the selected strategy followed by the post-processing of model response then executes it on the server, and displays the result.
• Leverage the infilling and commit-style of StarCoder to self repair the codes for potential errors
• Flexibility provided by allowing users to edit the generated code and re-run it for customized solutions.

• Became Kaggle Competition Expert
• Research work "OctoPack Instruction Tuning Code Large Language Models" is accepted as the SPOTLIGHT at ICLR 2024 (top 5%)
• Top 7% (Bronze medal) in Kaggle’s UBC Ovarian Cancer Subtype Classification and Outlier Detection (UBC-OCEAN) Competition

• My project Clothes Virtual Try On hit 275+ stars on GitHub! 🌟
• My collaborative research work, "StarCoder may the source be with you!" has been accepted at the TMLR (Transactions on Machine Learning Research).
• 700+ citations of my research work on Google Scholar
• My collaborative research work, "OctoPack Instruction Tuning Code Large Language Models" has been accepted at the Instruction Workshop @ NeurIPS 2023.
• Selected in Amazon ML Summer School 2023, Engaged in advanced ML modules and collaborated with leading Amazon ML Scientists.

• Secured Top 3 % global rank in Kaggle's 30 days ML challenge.