Applied Computer
Vision Engineer

I design and deploy production-grade computer vision systems for fixed industrial environments, converting raw detections into reliable process decisions using spatial and temporal logic. My work spans data curation, model training, post-processing logic, and real-time inference under strict operational constraints.

17K+ Images Trained
0.98 mAP50 Score
13+ it/s Inference
About

My work focuses on bridging the gap between object detection accuracy and real-world reliability. In industrial computer vision, high mAP alone is insufficient,systems must reason about position, sequence, and time.

I specialize in fixed-camera vision systems where spatial validation, temporal smoothing, and rule-based state tracking are used to transform frame-level detections into dependable operational signals. I take ownership of the full pipeline, from dataset curation and training to deployment, monitoring, and known failure analysis.

System Constraints

• Fixed camera placement and constant field-of-view
• Offline-only inference (no cloud access)
• Latency-sensitive pipeline under continuous operation
• No model retraining allowed post-deployment

Experience
AI Intern
Defect Scanner
JUL 2025 – JAN 2026
  • Problem: Frame-level object detections were insufficient to validate industrial drill-tightening operations reliably.
  • Approach: Trained YOLO-based models on 17K+ images (90K+ instances), then layered centroid-based spatial validation to confirm tool presence within predefined work zones.
  • Stability: Introduced temporal smoothing and state tracking to suppress transient false positives caused by motion blur and operator hand overlap.
  • Outcome: Achieved mAP50 of 0.98 with stable real-time inference at 13+ it/s, deployed via a Flask dashboard for continuous production monitoring.
DL Engineer Intern
FDAI
DEC 2024 – JUN 2025
  • Developed OCR pipelines using PaddleOCR and Tesseract to extract text, tables, figures, and links from diverse document formats.
  • Deployed offline LLMs locally via Ollama — no internet — for secure, low-latency inference.
Skills
Core Language
Python & AI Pipelines
Production-oriented workflows for real-time data handling.
Python Flask Streamlit
Computer Vision
Real-Time Object Detection
YOLO-based detection, OpenCV processing, ROI extraction for industrial use cases.
YOLO OpenCV ROI
Data & Training
Model Training & Curation
Custom dataset curation (17K+ images), data augmentation, CVAT labeling, centroid-based validation.
CVAT Augmentation mAP
Applied AI
Offline LLMs & OCR
Offline LLM inference via Ollama, OCR extraction with PaddleOCR and Tesseract.
Ollama PaddleOCR Tesseract
Data Management
Structured Data & Reproducibility
Experienced in handling structured datasets and ensuring reproducible workflows for experiments, data curation, and model training in industrial computer vision projects.
Data Curation Reproducibility Structured Data Data Handling
Collaboration
Version Control & Team Workflows
Proficient in Git and GitHub for team collaboration, including branching strategies, pull requests, and resolving merge conflicts to maintain clean, reliable code.
Git GitHub Branching PRs Merge Conflicts
Projects
📄
Document Extraction Tool
Built a Python tool using PyMuPDF to extract structured metadata from PDFs and validate values automatically. Developed a Streamlit app to generate structured Excel reports using OpenPyXL, reducing manual verification effort.
PyMuPDF Streamlit OpenPyXL Python
🔬
Production-Line Defect Scanner
Designed a fixed-camera industrial vision system where object detection alone was insufficient for decision-making. YOLO-based models were trained on 17K+ images (mAP50: 0.98), then augmented with ROI filtering and centroid-based spatial logic to validate drill-tightening actions.

To handle real-world noise, temporal smoothing and state-based validation were applied, reducing false triggers from partial occlusion and rapid operator motion. The system delivered consistent real-time performance at 13+ it/s through a Flask-based monitoring dashboard.
Design Considerations & Limits:
• Fixed camera geometry eliminated need for homography
• Centroid logic chosen over IoU (Intersection over Union) IoU = (Area of Overlap) / (Area of Union) due to tool shape variance
• Known failure case: extreme hand occlusion during fast motion
YOLO Flask OpenCV Python
Education
MSc. Information Technology
University of Madras, Chennai
Aug 2023 – May 2025