Apr 2023 – Mar 2024 · 11 mos · Karlsruhe, Baden-Württemberg, Germany · Remote

• Conducted research and implemented Adversarial Machine Learning techniques (FGSM, PGD) to evaluate and strengthen model robustness against adversarial attacks.
• Fine-tuned machine learning models using adversarially generated datasets to improve resilience and generalisation under attack scenarios.
• Executed model-stealing attacks (Copycat CNN) to identify vulnerabilities; developed and implemented protective algorithms such as model watermarking and fingerprinting.
• Performed validation and evaluation of fairness, toxicity, and bias in Large Language Models (LLMs), building frameworks to ensure ethical AI deployments.

May 2022 – Oct 2022 · 5 mos · Stuttgart, Baden-Württemberg, Germany · Hybrid

• at the Institut für Parallele und Verteilte Systeme (IPVS)
• Conducted research and implemented spatio-temporal word embedding models, enhancing Skip-Gram architectures to incorporate temporal and geographic context.
• Developed Python-based embedding pipelines and evaluation frameworks for temporal semantic shifts, regional dialect detection, and time-sensitive NLP tasks.

Jan 2021 – Oct 2021 · 9 mos · Bengaluru, Karnataka, India · Remote

• Developed a multimodal emotion recognition system, integrating BERT (text) and wav2vec (speech) embeddings into a unified transformer-based neural network architecture trained with a joint loss function.
• Implemented a Voice Activity Detection (VAD) pipeline using binary Convolutional Neural Networks and applied K-means clustering for effective Speaker Diarization.
• Achieved strong performance metrics with an Emotion Recognition F1-score of 0.88 and a Speaker Diarization Error Rate (DER) of 17.35%.

Mar 2019 – Mar 2020 · 1 yr · Gurgaon, India · On-site

• Designed and deployed end-to-end NLP pipelines using AWS container services for low-latency text processing and inference.
• Developed style transfer, language correction, data augmentation (21 modules), and personality detection models, boosting overall accuracy by 4.73%.
• Integrated classical NLP methods with BERT-based deep learning to enable context-aware rephrasing, achieving up to 0.93 semantic similarity.
• Presented strategic AI insights and emerging NLP innovations to leadership, directly shaping the product roadmap.

Jul 2018 – Dec 2018 · 5 mos · Gurgaon, India · On-site

• Architected and implemented a 15-layer attention-based encoder–decoder model for high-precision language translation, achieving a BLEU score of 37.13.
• Developed a domain-tailored BiLSTM-CRF model for pharmacological named-entity recognition in NIH protocol documents, reaching an F1 score of 0.83.
• Built ML pipelines with Keras, SpaCy, NumPy, Pandas, and Anaconda for reproducible, rapid integration of translation and NER into apps.

Jun 2017 – Jul 2017 · 1 mo · Agra Area, India

• This is an application fully developed in Codeigniter(PHP MVC). It was really a whopping useful application for MSME (Government of India) as it generates certificates automatically and have functions like direct CSV import and automatic QR code embedding on certificates for verification.

Jun 2016 – Jul 2016 · 1 mo · Agra Area, India

• In this internship I worked on ERP software for MSME(Govt. of India).In ERP software i worked on attendance module and courses module. It is fully based on Drupal, a PHP framework.