Hire Sravanthi R. - Fine-Tuning Engineer

* Spearheaded LLM fine-tuning pipeline on AWS, improving enterprise model inference accuracy by 42%, reducing retraining * latency by 35%, significantly enhancing customer GenAI solution adoption. * Optimized real-time media personalization and recommendation workflows, increasing user engagement by 38% and ad * revenue by 26%, directly contributing to measurable business growth metrics. * Implemented RAG retrieval pipelines for multi-terabyte datasets, boosting query response performance by 48% and * accelerating knowledge delivery for enterprise AI applications across diverse industries. * Designed and deployed large-scale LLM inference endpoints on AWS SageMaker, leveraging DeepSpeed and PEFT for * efficient fine-tuning and GPU utilization across multi-tenant enterprise environments. * Built streaming data pipelines using Apache Kafka and PySpark to feed real-time user interactions into embedding-based * recommendation and personalization layers for GenAI applications. * Integrated vector search and retrieval systems with FAISS and custom embedding pipelines, enabling low-latency RAG * responses and scalable AI-driven content recommendations. * Developed model evaluation frameworks using MLflow, custom metrics, and automated safety checks to monitor * hallucinations, bias, and drift across multiple fine-tuned foundation models. * Implemented Python SDKs and REST APIs to support enterprise consumption of GenAI services, enabling seamless * integration with existing applications and internal tooling on AWS. * Architected GPU cluster autoscaling and container orchestration with AWS EKS, Docker, and Kubernetes, ensuring reliable * real-time model serving and optimized resource utilization. * Built end-to-end media preprocessing pipelines using OpenCV and FFmpeg, integrating annotation workflows through Scale * Rapid, enabling high-quality datasets for supervised and generative AI tasks.

* Spearheaded Llama 2 fine-tuning pipelines on AWS, improving inference throughput by 42% and reducing first-token * latency across enterprise applications, enabling scalable real-time AI-driven content personalization. * Optimized RAG pipelines integrating Llama 2 with media embeddings, increasing customer engagement predictions by 38% * and accelerating e-commerce ad relevance, contributing directly to revenue growth and ROI metrics. * Designed token-level caching and dynamic batching in Llama 2 inference, lowering GPU costs by 27% while maintaining * model accuracy, supporting Meta’s large-scale AdTech and personalization workloads. * Engineered distributed Llama 2 inference architecture on AWS GPU clusters using PyTorch 2, DeepSpeed-MII, and Triton, * ensuring low-latency responses for high-demand real-time ML applications. * Developed multimodal media pipelines combining SAM and ImageBind embeddings with Llama 2 on AWS, enabling scalable * text-image/video integration for advanced AI personalization solutions. * Built vector search pipelines using FAISS integrated with Llama 2 on AWS, supporting retrieval-augmented generation, * semantic search, and cross-modal recommendation systems for e-commerce and content applications. * Implemented quantization and LoRA-based fine-tuning for Llama 2 models, leveraging FlashAttention2 and vLLM * frameworks to optimize inference performance across AWS GPU workloads. * Orchestrated real-time streaming data pipelines via Kafka and AWS S3, integrating Llama 2 with media and transactional * inputs for personalized recommendations and AdTech scoring systems. * Designed monitoring and observability stack on AWS using Prometheus, Grafana, and OpenTelemetry to track Llama 2 * inference performance, model latency, and system resource utilization across production environments.

* Improved real-time object detection accuracy by 23% on edge devices, enabling visually impaired users to safely navigate * urban environments with low-latency AI-driven audio guidance. * Optimized OCR and NLP pipelines, increasing scene-text recognition efficiency by 28%, accelerating narration generation, * and significantly enhancing accessibility experience for over 100 pilot users. * Deployed AI personalization models on AWS cloud, reducing response latency by 35% and adapting speech and object * detection dynamically to individual user preferences in real time. * Developed and deployed PyTorch and TensorFlow computer vision models, leveraging TFLite for mobile edge inference and * AWS SageMaker endpoints for large-scale scene understanding. * Engineered asynchronous media streaming pipelines using Python asyncio and WebRTC, integrating OCR, CV, and TTS * modules to deliver real-time multimodal accessibility feedback. * Implemented NLP/NLU and speech synthesis layers using HuggingFace Transformers, spaCy, and AWS Polly, enabling * personalized audio narration with low-latency real-time performance. * Built CI/CD pipelines and model versioning with Docker, Kubernetes, and AWS Lambda, ensuring continuous integration of * AI models, automated deployment, and robust monitoring of inference workflows.