Anika Shrivastava

Developed and optimized Natural Language Understanding (NLU) modules for chatbots, focusing on improving the accuracy of intent classification and real-time query interpretation.

• Engineered a BERT-based NLU module for chatbots, leveraging Hugging Face Transformers and scikit-learn to enhance named entity recognition and contextual intent analysis.
• Enabled accurate interpretation of user queries in real-time conversations through advanced NLU model development.
• Achieved an 83% intent classification accuracy on an internally curated test set by fine-tuning BERT and implementing robust input preprocessing techniques.

Designed and implemented automated pipelines for satellite telemetry data verification and anomaly detection, significantly reducing manual effort and improving data transmission reliability.

• Developed an automated pipeline for the verification of satellite telemetry (TM) PROM data and generation of structured test error reports.
• Minimized manpower and manual error by 30% across 5 ground stations by implementing a supervised Machine Learning approach.
• Implemented an anomaly detection pipeline leveraging the Random Forest algorithm, trained on labeled telemetry logs, to flag inconsistent or faulty data transmissions.
• Achieved over 95% accuracy in detecting anomalies in satellite telemetry data, enhancing data integrity.

Investigating latent spaces of autoencoders and developing a novel deep-learning solution for blind image deconvolution, aiming for robust reconstructions under diverse noise conditions.

Designed and trained a Multi-CNN Autoencoder model to reconstruct occluded regions and denoise images, achieving significant PSNR improvements without relying on clean-noisy image pairs.

Developed a real-time drowsiness detection system utilizing CNNs and OpenCV for high-accuracy eye state classification and accident prevention.

Authored a research paper characterizing Convolutional Autoencoder (CAE) and Denoising Autoencoder (DAE) latent manifolds as stratified manifolds, and Variational Autoencoder (VAE) as smooth product manifolds of symmetric positive definite and semi-definite matrix manifolds. Selected for oral presentation at the International Conference on Computer Vision Theory and Applications in Porto, Portugal, with an acceptance rate of 43.8%. Collaborative work with Dr. Samar Agnihotri and Dr. Renu Rameshan from Vehant Technologies Pvt. Ltd.