Ron Jailall

Driving end-to-end AI product development, focusing on Search, RAG, and Computer Vision.

• Search Relevance & Recommendation Systems: Designed and implemented real-time RAG-based inference pipelines to power dynamic, personalized product recommendations, directly optimizing for relevance and user engagement. Developed production-grade embedding model servers powering real-time search and recommendation web apps, ensuring low-latency retrieval.
• End-to-End Deep Learning (Matte Model Project): Took full ownership of a custom human matting solution, architecting a MobileNetV2-based model in TensorFlow 2/Keras to replace legacy SDKs. Managed the entire lifecycle: dataset curation (P3M-10k), custom data augmentation pipelines, training, and ONNX export for cross-platform deployment.
• Transformers & Optimization: Migrated Nvidia Riva/Triton microservices to AWS, optimizing transformer-based ASR/NLP models for scalability. Accelerated Stable Diffusion models for real-time generation, implementing latency optimizations for production video pipelines.

Led a cross-functional RAG team in a fast-paced startup environment, evaluating and deploying LLMs.

• LLM Fine-Tuning & Application: Collaborated with researchers to evaluate new prompting techniques and model architectures (Llama, Mistral) for customer-specific use cases. Modified C/C++ inference engines (llama.cpp) to support specific quantization and performance requirements for production APIs.
• Rapid Prototyping: Built React-based demos powered by LLMs and diffusion models to quickly validate product concepts with stakeholders. Created CI/CD pipelines (Terraform/AWS) to deploy ML inference servers, enabling the team to move fast from experiment to deployed endpoint.

Progressed from Engineer to Lead, introducing Deep Learning and Neural Search to the organization.

• Search & NLP: Prototyped neural search of video archives using segmentation and classification techniques (PyTorch, TensorFlow), significantly improving content discoverability. Created a C#/.NET utility to build sentence corpuses using TF-IDF and stemming, preparing data for IBM Watson speech recognition training.
• Model Training & Fine-Tuning: Developed a prefix-tuning dataset for the Nvidia Megatron LLM, gaining early experience in parameter-efficient fine-tuning (PEFT) methods. Utilized AWS Sagemaker and PyTorch to build and modify U-Nets for image analysis and segmentation tasks.
• System Ownership: Designed and built an AWS cloud-native Archive utility (Lambda, Batch) that became a core revenue-generating product, demonstrating ability to solve ambiguous problems end-to-end.