Himanshu Singh

At mytender, I spent 2 months building some pretty cool stuff. Think RAG pipelines optimized with LangChain to push retrieval precision, scalable frontend systems in Next.js stabilizing AI modules, and agentic orchestration refinements raising efficiency by 35%, the kind of work that keeps you up at night because it's actually interesting.

Here's where it gets interesting I engineered RAG pipelines, blending vector and keyword search for smarter retrieval, orchestrated multi-agent chains for seamless routing, and built a modular Next.js frontend that kept everything stable and responsive as new features shipped.

At TEXMiN, I developed a traffic detection model quantized to FP16, achieving 0.92 mAP@0.5 and 30–35 FPS on constrained hardware. I trained a custom YOLOv8l model, reduced its size by 42% with hybrid optimization, and fine-tuned parameters for quantization, pruning, and inference engines to maximize performance across Jetson Nano and Raspberry Pi 4 deployments.

Here's where it gets interesting I built and optimized quantized YOLO models, for real time edge deployment, blending advanced quantization and pruning to squeeze high performance from limited hardware. This work pushed the boundaries of what's possible with resource-constrained devices, making robust traffic analysis and mining equipment detection feasible in real world, low power environments.

At IIT Bhilai, I implemented advanced image forgery detection using TensorFlow, boosting classification accuracy by 15% through rigorous training and structured experiments. I optimized algorithms and streamlined workflows, cutting processing time by nearly 30%, and managed over 1,000 images organizing, cleaning, augmenting, and labeling datasets for robust training.

Here's where it gets interesting I implemented systematic preprocessing and annotation pipelines that boosted throughput by 30%, making large scale image forgery detection faster and more reliable. This approach not only improved accuracy but also dramatically reduced computation time, enabling scalable and efficient deployment for real world image analysis tasks.