Taya M.

– Content Generation API ∗ Designed and deployed highly scalable FastAPI endpoints to generate 20k+ product sheets (titles, meta descriptions, descriptions) in under 24 hours, reducing manual workload by 90%. ∗ Built a Streamlit dashboard to monitor client-specific content generation throughput, costs, and bottlenecks, enabling data-driven resource allocation. ∗ Integrated Grafana dashboards for trace and log monitoring, improving incident response time by 40%. ∗ Developed a rule-based scoring model to validate content compliance with client requirements, ensuring 95% adherence rate. ∗ Trained a probabilistic hallucination detection model to flag unreliable generations, reducing error rate in production. ∗ Automated deployment of APIs to Azure via GitLab CI/CD, improving release velocity. – Categorization API ∗ Developed FastAPI endpoints for taxonomy-agnostic e-commerce product classification, achieving 80% accuracy on unseen data. ∗ Trained a custom BERT model on internal large-scale product data for embeddings, combined with LLM inference for final predictions. ∗ Built a Streamlit app for Customer Success teams to track categorized products, clients, and associated costs, streamlining operations. ∗ Optimized inference pipeline to classify a product in <3 seconds. ∗ Deployed production-ready API on Azure using GitLab CI/CD pipelines. – LLM Reverse Proxy ∗ Designed a scalable FastAPI reverse proxy to manage multi-provider LLM calls (OpenAI, Anthropic, Gemini), supporting 1k+ requests/sec. ∗ Implemented auto-scaling mechanisms and centralized monitoring to optimize system reliability and performance. ∗ Developed a cost analytics dashboard, reducing provider costs by 20% through usage insights. ∗ Stored LLM calls and outputs in PostgreSQL, enabling fine-tuning of smaller models (LLaMA, Falcon) using Azure ML Studio, improving efficiency and lowering inference costs.

– Began as a Math/Statistics tutor, improving client communication and problem-solving skills through one-on-one teaching. – Built and deployed time series forecasting models (ARIMA, Prophet) for PhD researchers in physics, enabling accurate predictions and data-driven insights. – Delivered project-based AI solutions for startups in Europe and the Middle East, covering NLP, computer vision, and predictive analytics use cases. – Designed and implemented end-to-end solutions including data preprocessing, feature engineering, model training, deployment, and visualization. – Completed projects such as: ∗ Brain Tumor Segmentation using CNNs on medical imaging data. ∗ Insurance Pricing Models with regression and risk scoring. ∗ Retail Product Classification with LLM fine-tuning (achieved 85% accuracy on unseen data). ∗ Cybersecurity Data Dashboard with Power BI for real-time threat insights.

– Developed a Named Entity Recognition (NER) system to extract structured insights (victim, location, crime details) from unstructured police reports in Portuguese, reducing manual effort for law enforcement. – Implemented deep learning models leveraging Transformers for sequence labeling, achieving robust performance on complex text inputs. – Built an interactive platform with modules for model monitoring (KPI dashboards), human-in-the-loop corrections, and data import/export to support non-technical police officers and data scientists. – Designed workflows for model training, hyperparameter tuning, and evaluation with balanced accuracy metrics, improving reproducibility and transparency.