Ema – Python, OpenAI, LLM

An AI and cloud consulting initiative focused on helping organizations identify, evaluate, and implement AI-driven solutions. The work included designing scalable AWS-based architectures, defining technical roadmaps, assessing emerging AI technologies and data platforms, developing proofs-of-concept and MVPs, and enabling organizations to leverage data and AI more effectively across software development, analytics, and operational workflows.

• Lead a team of 4 people to create a RAG system over a matrix of current use cases for AI Native and AI Assisted SDLC for internal management;
• Developed AI-powered tools to support accessible software development by integrating LLMs into developer workflows;
• Designed and implemented backend services and RAG for chatbot and MCP solutions, ensuring scalability and reliability;
• Lead a team of 6 people to prepare an MVP for an IoT client with a lot of data and little idea on how to leverage it;
• Prepared infrastructure documentation and contributed to cloud and DevOps planning aligned with enterprise-grade standards;
• Led R&D initiatives on cutting-edge AI technologies to improve accessibility and developer productivity;
• Lead a team of 4 people to identify 12 AI Native Data Engineering and BI tools, reducing them to 2 recommended tools for the client;
• Built an email classifier app fully hosted in AWS to showcase the power of AI coding tools.

An AI-driven software development platform focused on improving accessibility and developer productivity through the integration of Large Language Models (LLMs) into engineering workflows. The project involved building scalable AI-powered applications, chatbots, and MCP-based solutions, conducting research and experimentation with emerging AI technologies, and designing cloud-native architectures to support enterprise-grade deployment and operations.

• Developed AI-powered tools to support accessible software development by integrating LLMs into developer workflows;
• Designed and implemented backend services for chatbot and MCP solutions, ensuring scalability and reliability;
• Prepared infrastructure documentation and contributed to cloud and DevOps planning aligned with enterprise-grade standards;
• Led R&D initiatives on AI technologies to improve accessibility and developer productivity.

Bioprocess simulation platform for a pharmaceutical company.

• Led AI and ML strategy for a next-generation bioprocess simulation platform for a biopharmaceutical company;
• Defined ML use cases, including sensitivity analysis, automated retraining strategies, and model lifecycle management;
• Collaborated with cross-functional teams to align technical ML solutions with bioprocessing domain requirements and regulatory expectations;
• Proposed AI-powered assistant features for contextual guidance, experiment design, and performance troubleshooting.

AI Assistant and retrieval-based knowledge product used by 1000+ monthly active users.

• Developed and scaled an AI Assistant to 1000+ monthly active users;
• Led the implementation of an evaluation system for the RAG pipeline;
• Built RAG workflows for large-document ingestion and querying with Pinecone, OpenAI embeddings, and models;
• Implemented LLM API integrations for narrative generation and chatbot building using OpenAI, Gemini, Gemma, and Anthropic Claude.
• Led topic modeling of chats;
• Prototyped agent-based evaluation and brainstorming systems with LangChain;
• Applied MLOps and LLMOps practices with Terraform, GitHub Actions, and Docker.

An AI-powered storytelling platform for children that generates personalized illustrated storybooks using generative AI. The project combined image generation technologies, including Stable Diffusion, ControlNet, and LoRA fine-tuning, to create consistent characters and visual narratives, while leveraging multiple large language models to generate engaging story content. The solution explored and evaluated various AI approaches to deliver a high-quality, interactive storytelling experience and was recognized at a major global developer conference.

• Co-led development an AI-generated storybook product for children that was selected and featured at Google I/O 2023 in San Francisco;
• Designed and implemented a character consistency pipeline combining ControlNet, Stable Diffusion, and LoRA weights after evaluating multiple image generation approaches;
• Integrated narrative generation using OpenAI, Gemini, Anthropic, and Mistral APIs to produce coherent, age-appropriate children's stories;
• Researched and benchmarked image generation methods to identify the optimal approach for maintaining visual character consistency across storybook pages;
• Led prompt engineering and model selection for multi-LLM narrative generation;
• Collaborated cross-functionally to ship a production-grade generative AI product to a public audience at a major industry event.

An end-to-end machine learning project focused on building an image classification pipeline for identifying over 40 types of construction machinery. The system was designed, trained, and optimized using deep learning models (ResNet34, ResNet50, ResNet101) and later deployed as a mobile application for field use. The work included data preprocessing, dataset cleaning, model training, and evaluation using a combination of Python-based ML frameworks. The solution enabled practical on-site machine recognition through an Android application.

• Built a multiclass object detection pipeline for counting and classifying microplastic particles collected at the beaches of the Canary Islands;
• Achieved an overall F1 score of 96.41% on the object detection task;
• Worked with YOLOv5 for object detection, UNET for semantic segmentation, and VGG16 for classification;
• Achieved results 3.6x better than the state-of-the-art model on precision, recall, and F1.

An NLP-focused machine learning project centered on fine-tuning transformer-based models for text classification tasks. The work involved adapting BERT and GPT-2 models for both multiclass and binary classification across two datasets, including data preparation, model training, and evaluation. The solution achieved high performance in classification accuracy and F1 score, and contributed to an academic research publication presented at the SwissText conference.

• Fine-tuned multiclass and binary language models, including BERT and GPT-2, for text classification on two datasets;
• Achieved 99% accuracy and 98% F1 in text classification tasks;
• Co-authored a research paper presented at the SwissText conference.

Classification pipeline for 40+ construction machines that was later turned into an Android app.

• Built the ML backend of a classification pipeline for more than 40 construction machines for a client;
• Performed data preprocessing and cleaning;
• Contributed to a solution that was later turned into an Android app.