Post Job Free
Sign in

Scalable Computer Vision & NLP Engineer

Location:
New Delhi, Delhi, India
Posted:
October 23, 2026

Contact this candidate

Resume:

RAHUL GUPTA

Gurugram, Haryana, India ***************@*****.*** 965******* LinkedIn GitHub Portfolio PROFESSIONAL SUMMARY

Computer Vision AI/ML Engineer with 1.6+ years of experience researching, building and deploying scalable real-time CV systems across Object-detection, Image-segmentation, Multi-Object-tracking. Proficient in CNN, YOLO, Transformer-based architectures with strong expertise in production multi-camera edge-inference optimization using TensorRT, ONNX and NVIDIA DeepStream. Familiar with CUDA/GPU programming, Linux setup and AWS cloud. Proficient in managing, maintaining and debugging the applications in production.

EXPERIENCE

Software Developer (AI/ML) Jan 2026 – Present

Galactic Client Services — Gurugram, Haryana

Developed and optimized end-to-end Computer Vision training and scalable inference pipelines of multi-object tracking (MOT), detection, segmentation using both deep learning and computer vision techniques. Managed multiple RTSP streams with low- latency and cost optimization at good accuracy.

Monitoring, Managing and Debugging the end to end pipelines through logs. Adding features, fixing the bugs and handling escalation of the client.

Developed speech-to-text and speech-to-speech conversational AI voice bots integrated with WebRTC real-time calling systems using low-latency streaming pipelines including Voice Activity Detection (VAD) and LLM-based conversational orchestration. Data Process Associate June 2025 – Jan 2026

IGT Solutions (Atain) — Gurugram, Haryana

Performed end-to-end data preprocessing for e-commerce industry including image processing, normalization, augmentation, annotations and transformation pipelines to ensure high-quality model-ready datasets at scale for image classification, object recognition, and text classification tasks.

SKILLS

Programming Languages Python, C++, SQL

Deep Learning CNN, YOLO, Transformer architectures & ViT, SAM, MOT, Quantizations, Pruning, Hyperparameter-tuning Computer Vision Camera Calibration, Streo-Vision, Geometry, Point-clouds, Image Processing Frameworks PyTorch, TensorFlow, Keras, OpenCV, Open3D, Detectron, Deepstream, MediaPipe, HuggingFace APIs FastAPI / RestAPI

MLOps GCP, Docker, CUDA/GPU, CI/CD, Linux, TensorBoard, MLflow, Grafana, W&B, ONNX, TensorRT Databases PostgreSQL, Vector Databases, Redis

PROJECTS

PPE (Personal Precaution Equipment) Compliance through CCTV Alerts July 2026 It detects whether the workers in the factory are wearing gloves or not and generates real-time alerts on CCTV streams or monitoring screens. It is classified into two parts — the training pipeline and the inference pipeline. The training pipeline consists of two computer vision and deep learning models and one tracking algorithm: 1. YOLOv11 for worker detection (initialized with COCO pretrained weights) 2. Fine-tune another YOLOv8 for obtaining the specific ROI (hands) — fine-tuned 3. Associate ByteTrack with the first detection model to get ReIDs In the inference pipeline, significant engineering was done to scale the system efficiently. The latency is approximately 130 ms end-to-end inference at 15 FPS. ONNX and NVIDIA TensorRT engines are used along with the DeepStream video streaming pipeline.

ByteTrack assigns unique IDs to workers, and temporal voting is used so that predictions are made based on the average of multiple frames. This helps reduce unstable predictions across consecutive frames. RTX 5080 GPU handles 10–12 concurrent CCTV streams with FP16 optimization. The models achieve approximately 87–90% overall precision and recall. Precision and minimizing false positives (FP) are the most important factors in this project, as the goal is to avoid false alarms. Fine-tuning Vehicle OCR Feb 2026

In this project, vehicle number plate transcription was performed from vehicle images using PaddleOCR. I fine-tuned two PaddleOCR models for this task:

1. PaddleOCR Det v4

2. PaddleOCR Rec v4

PaddleOCR Det v4 was used to detect the text regions from the license plate ROI, while PaddleOCR Rec v4 was used to recognize the characters and generate the final transcription output. The models were fine-tuned on approximately 1,500 manually annotated custom-labeled images containing different lighting conditions, plate formats, blur levels, and viewing angles to improve robustness and OCR accuracy.The pipeline generates the transcription from the extracted license plate region, and stores the structured outputs in JSON format with the assigned plate unique id for further processing and analytics. Achieved Recall of 89%, mAP50 of 92%, and Character Accuracy of 91% EDUCATION

Diploma — IIT Guwahati Jan 2026 – April 2026

In Computer Vision and Natural Language Processing Post Graduation — University of Delhi Apr 2023 – Apr 2025 Specialization in 'Cognitive Science' in the domain of Philosophy, which generally deals with the study of processing of Artificial Intelligence, Mind modularity and Cognition. It combines the aspects of Computer Science, Psychology and Neuroscience. Graduation — Central University of Allahabad Apr 2020 – Apr 2023 CERTIFICATIONS

OpenCV — OpenCV University July 2026 · Certificate AI on Jetson Nano — NVIDIA Jun 2026 · Certificate

Computer Vision & Natural Language Processing — IIT Guwahati & NPTEL

April 2026 · Certificate

Model Context Protocol(MCP) — Anthropic April 2026 · Certificate Data Science Masters(Machine Learning) — PW Skills March 2024 · Certificate



Contact this candidate