Alexandra B.

Developed proof-of-concept speech recognition system using MFCC feature extraction + DSP techniques (FFT, DCT-II, Mel filters) to reduce dimensionality and enable efficient ML model training. Key Responsibilities & Achievements: Clarified business needs w/ client/team; presented insights/research on DSP, speech math, SOTA architectures Built entropy-based VAD for noise-robust speech separation; framed signals w/ Hamming windowing FFT → Mel filter bank → DCT-II pipeline extracting decorrelated, perceptually-relevant MFCC coefficients Benchmarked SVM/RF/HMMs; trained 1D-CNNs on MFCCs; LSTM-RNNs modeling phonetic temporal dependencies Audio augmentation (pitch shift, time-stretch, Gaussian noise); hyperparameter optimization maximizing F1 across accents/noise Optimized Python pipeline (Librosa/NumPy); MLFlow experiment tracking + SageMaker training jobs Tech Stack: Python/NumPy/Pandas/Matplotlib, PyTorch/Hugging Face/Librosa, AWS (S3/EC2/SageMaker/ECR)Flask/PostgreSQL, MLflow

Developed AI software for ag robots covering pre-harvesting (RGB/NIR-based crops/weeds detection, disease spotting, field recognition) and harvesting (Sonar-driven fruit count/maturity estimation). Key Responsibilities & Achievements: Partnered closely with client to refine goals, addressing feasibility/end-use gaps for reliable, interpretable AI Built generic model quantization libraries and dataset staging pipelines—cut training time and eliminated duplication Benchmarked UNet, DeepLab, Mask R-CNN, ViT for RGB+NIR segmentation; automated pesticide/fertilizer application Adapted backbones for field classification, disease detection; implemented Multimodal YOLO/FusionNet fusion techniques Designed Sonar preprocessing + multi-output CNNs for fruit estimation; optimized models via OpenVINO for edge deployment Engineered real-time video pipelines for obstacle detection; added IG/XRAI interpretability layers Managed Linux robot filesystems; explored GAN/classic denoising; led code reviews

Built evidence-based decision tool using ML to analyze patents/trending news, uncovering profitable R&D directions for researchers. Key Responsibilities & Achievements: Refined scope via client/stakeholder comms; analyzed ops, led training on improvements/data support Classified docs (tf-idf/Word2Vec + Linear/SVM/RF); NER (RNNs) for orgs/institutions/names/scientific terms; extracted titles/years EDA/prep: Identified key doc sections; Topic Modeling (LDA/NMF/clustering); parsed news for entities/topics; trended analysis by year Populated vector DB w/ embeddings/entities/relations; custom semantic similarity for cross-category matching; built recos (patents ↔ news) Custom eval metrics; automated news collection (Bash/cron); Sagemaker labeling pipeline (AWS Lambda transforms) Doc typing (Doc2Vec/tf-idf); visualizations (Plotly); algo optimizations Tech Stack: Python/Linux/Bash, PyTorch/NLTK/Hugging Face/Gensim, Pandas/NumPy/Scikit,MLflow/Jupyter