s
shyam_ranasara

Shyam

@shyam_ranasara

Machine Learning Enigneer

Indien
Englisch
Einige Informationen werden in englischer Sprache angezeigt.
Über mich
Hi, I'm Shyam Ranasara, a Machine Learning Engineer & Full-Stack AI Developer specializing in Computer Vision, Real-Time MLOps, and Generative AI. I build production-grade AI systems: custom YOLOv8 detection, FaceNet facial recognition, Swin Transformer & FAISS vector search, 4-bit QLoRA LLM fine-tuning, and responsive web applications. Whether you need a real-time CCTV pipeline, an intelligent document parser, or a full-stack Python/React app, I deliver tested, containerized, and lightning-fast solutions. Let's bring your AI project to life! ... Mehr lesen

Kompetenzen

s
shyam_ranasara
Shyam
offline • 
Durchschnittliche Antwortzeit: 3 Stunden

Meine Dienstleistungen

Entwicklung von KI-Chatbots
I will build a custom rag ai chatbot for your pdfs, docs, and website using langchain
Computer Vision
I will train and deploy a custom yolov8 object detection model for images and video

Portfolio

Arbeitserfahrung

Indsac

Machine Learning Intern

Indsac • Vollzeit

Feb 2026 - Present • 9 mos

Engineered an AI-based surveillance system for real-time visitor monitoring and security automation, reducing manual security check load by ~60% across 4 active CCTV feeds. Built a face recognition pipeline using HuggingFace FaceNet embeddings and FAISS vector indexing, achieving sub-100ms identity matching for 500+ profiles. Developed a real-time video pipeline in OpenCV to detect, track, and capture snapshots at 25 fps, handling concurrent streams without frame drops. Designed an automated alert engine flagging unrecognized individuals within 3 seconds, integrated with SMS and email notifications. Developed a Flask-based web application for visitor approval, access logs, and system monitoring; containerized using Docker.

Suvidha_Foundation

Machine Learning Research Intern

Suvidha Foundation • Vollzeit

Dec 2025 - Feb 2026 • 2 mos

Researched multi-document summarization for Indian English using LLMs; surveyed 15+ architectures (T5, BART, PEGASUS, LLaMA) to inform model selection. Fine-tuned Llama 3-3B using QLoRA (4-bit) via HuggingFace PEFT library, achieving ROUGE-1 49.06, outperforming PEGASUS baseline by 6.2 points. Designed 4-bit quantization pipeline enabling large-model fine-tuning on a single T4 GPU, cutting memory usage ~40% vs full fine-tuning. Benchmarked LoRA vs QLoRA across speed, memory, and ROUGE metrics; QLoRA delivered 3x lower compute vs full fine-tuning; tracked experiments using MLflow.