Pedro – LLM, RAG, LangChain

An AI-powered simulation platform that enables users to create and run structured persona-based simulations through self-serve workflows. The product combines agentic LLM capabilities with voice simulation orchestration, integrating speech-to-text, LLM, and text-to-speech services to streamline simulation setup and support scalable customer usage.

• Build a self-serve simulation product with a structured, form-based persona workflow, reducing simulation creation time from 1–2 days to about 1 hour and increasing throughput from 1 to 10 simulations per day.
• Develop an agentic LLM pipeline that converts free-form user prompts into structured persona configurations, reducing setup time from about 1 hour to 40 seconds.
• Design and implement voice simulation orchestration across third-party and self-hosted components, integrating Speech-to-Text (STT), LLM, and Text-to-Speech (TTS), contributing to a 5x increase in simulation usage and customer adoption.

An enterprise AI and machine learning platform providing services for clustering, classification, vector embeddings, and text and image generation. The platform supports production-grade AI workloads for enterprise customers, with scalable model deployment and optimization capabilities.

• Led development of 20+ ML/AI services across clustering, classification, vector embeddings, and text/image generation, serving 20+ enterprise customers with Python, NumPy, Scikit-Learn, PyTorch, Hugging Face, ChatGPT, and Llama.
• Built and operationalized MLOps workflows from development to production using Docker, Kubernetes, AWS SageMaker, GCP Cloud Run, FastAPI, ONNX optimization, and model quantization, improving model speed by up to 30%.
• Delivered scalable AI solutions for enterprise use cases, supporting reliable deployment and production operation across diverse ML workloads.

A surveillance solution that integrates machine learning applications to support intelligent monitoring and analysis. The platform combines video surveillance capabilities with ML-powered functionality for more efficient and automated monitoring.

• Built NVIDIA GPU-accelerated Kubernetes infrastructure for computer vision workloads, with reusable Kubeflow components for training, inference, and monitoring.
• Developed reusable Kubeflow components for model training, inference, and monitoring.
• Deployed a hard-hat detection model on in-house infrastructure with approximately 95% accuracy.