cv
General Information
| Full Name | Divyanshu Kumar |
| Languages | English, Hindi, Bhojpuri, Bengali (limited proficiency), Marathi (very limited proficiency) |
Experience
-
2023 - Present ML Research Engineer
Enkrypt AI - Building a LLMOps platform that provides access control, monitoring, evaluation and guardrails for LLMs.
- Developed a comprehensive risk identification framework utilizing Red Teaming methodologies and implemented mitigation strategies through the establishment of guardrails within LLMs.
- Built a framework for Differentially Privacy enabled finetuning of LLMs utilizing Opacus and HuggingFace libraries. The framework enables training method like- Parameter Efficient Trainings(PEFT) as well as SFT with DP Gaurantees.
- Developed a propierty framework named SecureLLM that leverages Homomorphic Encryption to encrypt pretrained LLMs.
-
2022 - 2023 Research Associate
Indian Institute of Sciences, Bangalore, India - Discovered a unique approach to unlearning targeted features/classes of GAN models via perturbations in the latent space.
- Interpreted a novel method to improve generative model robustness with differential privacy. Conducted experiments with MNIST, CIFAR10, and CelebA datasets, evaluating accuracy (FID) and privacy preservation.
- Developed a knowledge-distilled version of mBART for Hindi to Kannada using Hindi to Sanskrit and Sanskrit to Kannada.
- Added an embedding layer to IndicBART's encoder to accommodate parts of speech (POS) and implemented a POS model to generate embeddings for compatibility with the input embedding.
-
2021-2022 Intern
AIcrowd - Developed baselines for NLP and Computer Vision puzzles using deep learning models such as LSTM, BERT, and ALBERT. Utilized libraries like OpenCV (CV2), YOLO, and Detectron for Computer Vision.
- Constructed challenging puzzles focusing on ML tasks, such as image classification and segmentation, and sentiment analysis, using GANs and transformers to help students and professionals.
Education
-
2022 B.Tech(ECE)
Narula Institute of Technology, Kolkata, India - CGPA: 8.78/10
Skills
- Programming Languages: Python, Golang, Rust, JavaScript, C++, Java
- Libraries: PyTorch, FastAPI, Llama-index, Langchain, TensorFlow
- Tools: Git, Docker, GitHub Actions, Vim, Latex
Open Source Contributions
-
2022 NimbleEdge - Google Summer of Docs(Mentor)
- Worked with the NimbelEdge's Technical, Content team and GSOD mentee to audit, prepare and develop technical documentation, tutorials, and case studies for EnvisEdge. Google Season of Docs provides an opportunity for technical writers worldwide to contribute to open-source software.
- Organised weekly community meetups for community members to engage new community members in the development of EnvisEdge by code, documentation, or design.
-
2020 - 2021 OpenMined - Mentor
- Upbring Mentees towards learning about OpenMined through a well-directed Udacity Course on Secure and Private AI using PySyft.
- Resolving Technical Doubts and guiding them to the right resources to make them even with the Community and bring meaningful contributions.
- Maintaining a centralized record on the progress of the Mentees and participate in Weekly Syncs to discuss the progress.
Academic Interests
-
Causality
-
Generative Models
-
NLP
-
Privacy Preserving AI
-
Recommender Systems