● Led end-to-end development of a multilingual LLM-powered translation system, expanding dataset size by 1,650% and improving BLEU score performance by 50%. ● Integrated AI-powered user suggestion and follow-up systems, boosting daily active users by 20% and increasing subscriptions by 25%. ● Reduced model training time by 80%, thereby reducing cloud costs by 40% via optimization of training pipelines on GCP. ● Built and deployed content moderation models using NLP classification and sentiment analysis, cutting moderation latency by 30%. Project: Yoruba Version LLM. ● Yoruba LLM – Transformer-based LLM trained for native language generation tasks. Managed full architecture, tokenizer prep, and evaluation metrics.

● Orchestrated a production-grade Kubeflow pipeline deployed on Google Vertex AI, automating the full LLM lifecycle, including data loading from GCS, preprocessing, model training, evaluation, and conditional deployment. ● Used Cloud Build, Docker, and Vertex Pipelines to enable reproducible and automated LLM workflows in a cloud environment. ● Implemented automated decision logic to deploy models only when evaluation metrics surpass defined thresholds.

● Curated and quality-checked large-scale training datasets, enhancing accuracy in generative LLM systems. ● Evaluated model performance using cross-validation and offline metrics, identifying bias and adjusting labeling schemes. ● Collaborated in shaping AI roadmap for multilingual NLP products, integrating ethical considerations and responsible model usage practices.

● Fine-tuned LLMs (Mistral, Llama2) for text summarization and content generation using PEFT techniques, reducing fine-tuning compute costs by 40% while improving text coherence by 25% (measured by ROUGE scores), leading to higher user engagement and content adoption rates. ● Built an AI-powered document retrieval system with LangChain, OpenAI GPT, and vector embeddings, reducing manual search time by 60%, which enhanced workflow efficiency and cut operational costs by minimizing time spent on information retrieval. ● Collaborated with co-engineers on best practices of NLP model deployment, improving team proficiency in fine-tuning and productionizing AI models, leading to faster model iteration cycles and more scalable AI solutions, ultimately reducing development costs by 10%.