Joseph Rapaka

• Led a team of five engineers to architect and build an agentic AI chatbot powered by a LangGraph-based multi-agent architecture, orchestrating multiple background agents to autonomously generate multi-scene storyboards, videos, and comic-style visual narratives using Gemini multimodal models.
• Built enterprise Retrieval-Augmented Generation (RAG) pipelines leveraging LLMs, vector embeddings, and hybrid search to transform unstructured corporate documents into structured brand knowledge, audience insights, and creative guidelines, enabling context-grounded generation of creative concepts and ad campaigns.

Client: Mondelez International

• AIDA Platform Development: Contributed to the development of the AIDA Platform to streamline campaign asset creation and concepting. Enabled end-to-end workflows covering ideation, connection, and automated asset generation for launching campaigns and new product initiatives across multiple brands.
• Image Generation & Model Fine-tuning: Fine-tuned Bria Vision Models (Bria 2.3, 3.2, 4B), Imagen 1 to improve brand-specific accuracy for global brands such as Milka, Cadbury, Chips Ahoy, and Lacta. Integrated user-provided concepts with structured knowledge bases (Creative, Brand, Audience, Channel Insights) to generate engaging, brand-aligned, and audience-tailored ad scripts.
• Vision Capabilities in AIDA: Designed and Developed Image editing features such as Product Placement, OutPainting, Inpainting, Image Generation that can be utilized in AIDA application.
• Implemented RAG (Retrieval-Augmented Generation) pipelines using LLMs to extract structured brand information, audience segmentation insights, and creative guidelines from unstructured corporate documents, improving data-driven decision-making and generate a Creative Concepts and Ads campaigns.

• Client: Debeers Ltd
• Developed and implemented Deep learning Yolo models for object recognition and tracking, resulting in 90% increase in accuracy compared to previous models.
• Designed and maintained application called CVRMS for risk assessments on live camera feed and maintaining safety protocols in mining industry.
• Designed and Implemented multiprocessing techniques to efficiently handle simultaneous data processing from multiple camera feeds.
• Developed various object tracking methods for mining applications, capable of maintaining tracking accuracy even when camera angles are changed.

MLOps

• Constructed machine learning pipelines encompassing model labeling, training, and deployment processes utilizing AWS SageMaker and ECR (Elastic Container Registry).
• Designed Terraform scripts to automate AWS infrastructure setup for seamless service deployment, including EC2 instances, SageMaker, Lambdas, S3 buckets, and IAM roles.
• Developed automated testing scripts to assess performance of trained models, ensuring accuracy and reliability across diverse datasets and scenarios.
• Acted as team leader in group projects, delegating tasks and providing feedback
• Contributed to open-source AI projects as a way to stay current with industry advancements and give back to the community.

• Client: AFLAC Insurance
• Developed NLP-based application for automating insurance claim processing, specifically focusing on document analysis utilizing Named Entity Recognition (NER) techniques using BERT models.
• Engineered robust text classification model capable of accurately categorizing over 85 different types of documents resulting decrease in 20% processing time and improved application performance by 50%.
• Developed automated Python scripts to seamlessly transfer application outputs to AWS RDS databases. Integrated these results into application for streamlined access and presentation to end-users.
• Worked closely with domain experts to ensure accurate representation of problem context within developed models, enhancing their real-world applicability

• Constructed object detection models such as SSD and SSD Lite utilizing transfer learning techniques, leveraging multi-GPU training for accelerated performance and improved accuracy.
• Quantized trained model weights into ONNX format and optimized for TensorRT to assess improvements in model inference speeds.