1. AI Research Intern — FORVIA HELLA (AHEAD Project)

   Lippstadt, Germany Aug 2023 – Feb 2024

   Developed perception models for adaptive headlights using stereo vision and deep learning. Improved night-time detection accuracy by 30% and achieved real-time 3D localization with embedded GPUs.

   ▼

   

   

   

   

   

   ❮ ❯

   Goal

   Improve night-time perception for adaptive headlights by enabling reliable object detection and lane understanding under glare, rain, and low-illumination conditions.

   Approach

   • Adapted and fine-tuned YOLOv8 for glare-robust detection using extensive augmentation and hyper-parameter tuning.
   • Designed a modular perception pipeline combining detection, segmentation, and multi-object tracking (DeepSORT + OpenCV).
   • Implemented stereo triangulation and calibration to obtain real-world 3D localization and lane geometry.
   • Performed robustness and uncertainty analyses for varying lighting, motion blur, and sensor noise.
   • Collaborated with optics and embedded teams to validate models in prototype headlight systems.

   Results

   • Improved low-light detection mAP by ≈ 30% compared to baseline.
   • Achieved highly stable 3D localization validated against laser ground truth.
   • Enabled real-time inference through Dockerized GPU pipelines for reproducible testing.

   This work was presented to the R&D and perception teams at HELLA and influenced ongoing ADAS prototype development.

   Tech Stack: Python · PyTorch · OpenCV · YOLOv8 · DeepSORT · ONNX Runtime · Docker · Stereo Calibration · Triangulation

2. Master’s Thesis — Stereo Vision for Adaptive Headlight Systems

   Lippstadt, Germany Mar 2024 – Nov 2024

   Built a lightweight stereo CNN for long-range depth estimation and 3D object localization. Generated 9,000+ synthetic pairs in Unreal Engine 5 and optimized inference latency to 70 ms.

   ▼

   

   

   

   

   

   ❮ ❯

   Goal

   Develop a lightweight deep-learning model for dense depth estimation and 3D object localization in adaptive headlight control.

   Approach

   • Designed a stereo CNN (autoencoder + cost-volume fusion) for real-time depth perception.
   • Generated ≈ 9,000 synthetic stereo pairs in Unreal Engine 5 with calibrated intrinsics to simulate weather, glare, and reflections.
   • Applied domain randomization and fine-tuned on real test data to minimize the sim-to-real gap.
   • Combined depth maps with YOLO detections for object localization and trajectory estimation.
   • Exported optimized ONNX models to embedded GPUs and validated inference performance jointly with hardware and optics teams.

   Results

   • Achieved 3% D1-all error on KITTI benchmark and ≈ 95% depth accuracy at 30m on real data.
   • Reduced inference latency from 120 → 70 ms through ONNX Runtime optimization.
   • Delivered complete calibration, benchmarking, and uncertainty documentation adopted by HELLA’s internal R&D teams.

   This work was presented to the R&D and perception teams at HELLA and influenced ongoing ADAS prototype development.

   Tech Stack: PyTorch · YOLOv8 · Unreal Engine 5 · OpenCV · CUDA · ONNX Runtime · Docker · Calibration QA · Synthetic Data Generation

3. Automation Engineer — Indpro Electronic Systems Private Limited

   Pune, India May 2019 – Mar 2020

   ▼

   • Developed and deployed PLC control logic (ABB AC800M) for sugar-industry automation systems.
   • Designed HMI dashboards and supported on-site commissioning and troubleshooting across multiple facilities.

   Tech: ABB AC800M PLC, PLC Programming, HMI (SCADA), Industrial Automation

Browse My Recent

Projects

Custom YOLO for Robotic Bin Picking

Developed a custom YOLO model for detecting objects in bin picking tasks, achieving 98% precision in 6D pose estimation.

TensorFlow YOLO Bin Picking Blender OpenCV Matplotlib

• Developed a YOLO-based 6D pose estimation model in TensorFlow for precise object detection and orientation prediction in bin-picking tasks.
• Generated and annotated hybrid datasets using Blender (synthetic) and real camera images to improve robustness under varying lighting and occlusion.
• Integrated the perception system with an ABB YuMi cobot for automated pick-and-place; achieved 98% grasp success rate in real industrial tests.
• Optimized inference speed and detection stability for edge deployment, ensuring reliable operation in dynamic environments.

Github More Details

Deep Stereo Vision for Nighttime Driving Scenes

Built a lightweight stereo CNN for adaptive headlights — achieving 95% depth accuracy at 30m and 3% D1-all on KITTI.

PyTorch Unreal Engine 5 MATLAB OpenCV Stereo Depth

• Developed a lightweight stereo CNN in PyTorch for real-time depth estimation in low-light automotive environments.
• Generated 9,000+ synthetic stereo pairs in Unreal Engine 5 simulating illumination, reflection, and weather variations to enhance generalization.
• Achieved 3% D1-all error on KITTI and 95% real-world depth accuracy at 30m; optimized inference latency from 120 → 70ms via ONNX Runtime.

Github More Details

3D Reconstruction with Gaussian Splatting

Converted monocular smartphone video into a dense 3D scene using COLMAP preprocessing and Gaussian Splatting rendering.

COLMAP WSL Supersplat Gaussian Splatting Open3D

• Extracted camera poses and sparse point clouds using COLMAP from monocular input video.
• Used Gaussian Splatting to render photorealistic 3D views with transparency and shading effects.
• Visualized and explored outputs in Open3D, and investigated mesh conversion workflows.
• Handled lighting, occlusion, and splat shaders for high-fidelity rendering of complex indoor scenes.

Github More Details

Object Instance segmentation with FastAPI and AWS

YOLOv11 instance segmentation deployed via FastAPI, containerized with Docker on AWS EC2/ECR, and automated using GitHub Actions CI/CD.

YOLO AWS Docker CI/CD FastAPI Segmentation

• Real-Time Object Detection & Segmentation: Uses YOLOv11 for accurate instance segmentation.
• FastAPI Backend: Serves the model predictions via REST API.
• User-Friendly Web Interface: Allows users to upload images and view detected objects.
• Dockerized & Cloud-Hosted: Containerized using Docker, pushed to AWS ECR, and deployed on AWS EC2.
• Automated CI/CD Pipeline: Uses GitHub Actions for continuous deployment.
(live demo may not work because of EC2 instance being overused)

Live Demo

Github More Details

Anomaly Detection for PCB Inspection

Built a lightweight visual anomaly detection system using PaDiM and ONNX for industrial defect detection.

ONNX Runtime Anomaly Detection ONNX Matplotlib MobileNetV3

• Extracted multi-level features from MobileNetV3 and computed Mahalanobis-based anomaly scores at pixel level.
• Exported the feature extractor to ONNX and performed inference using ONNX Runtime for cross-platform compatibility.
• Visualized anomalies with heatmaps and binary masks using images from the MVTec AD dataset.

Github More Details

U-Net for Biological Image Segmentation

Trained a U-Net model on fluorescence microscopy images to segment nuclei structures with high accuracy.

Colab PyTorch Segmentation Matplotlib

• Parsed and preprocessed instance masks into binary labels using CLAHE and Gaussian blur.
• Built a lightweight U-Net in PyTorch; trained for 100 epochs on nuclei segmentation dataset.
• Achieved Dice score of 0.89 and IoU of 0.82 on the validation set.
• Visualized predicted masks and prepared inference-ready model using Colab.

Github More Details

Conversational AI Chatbot with RAG & LLM Fine-Tuning

Developed a GPT-based conversational AI assistant, integrated FAISS vector search, and deployed using Docker & Hugging Face Spaces.

Hugging Face / Transformers FAISS RAG

Github Live Demo

GaugeVision

The goal is to simplify gauge monitoring and tagging by using YOLOv11 pose model and QR detector.

YOLOv11 OpenCV RoboFlow

• Trained a custom YOLOv11 pose model to detect keypoints on analog pressure gauges.
• Integrated OpenCV's QR code detector to read identification tags placed on or near gauges.
• Combined both outputs visually to show needle direction and QR data in one intuitive display.

Github More Details

Edge Detector GUI (Qt + OpenCV + C++)

Built an interactive desktop app using Qt and OpenCV for visualizing Sobel and Canny edge detection with real-time threshold sliders.

OpenCV QT C++

• Integrated OpenCV with Qt to enable real-time image processing using Sobel and Canny filters.
• Added dynamic UI with sliders and dropdowns for live threshold tuning and filter selection.
• Designed clean modular C++ logic with CMake for cross-platform build compatibility.

Github More Details

BLIP-1 Optimization for Efficient Image Captioning

Optimized BLIP-1 image-captioning on CPU via dynamic quantization and pruning, achieving 1.42× speedup with no accuracy loss.

Quantization L1 pruning

• Baseline BLIP-1 CPU inference in 0.69 s for correct captions.
• Applied dynamic quantization to all Linear layers, reducing latency to 0.48 s.
• Combined 30% L1 pruning + quantization for a 1.40× speedup.
• Preserved caption quality (“a man and his dog”) across all variants.

GitHub More Details

Stereo Point Cloud Reconstruction

Built a stereo vision pipeline for 3D point cloud reconstruction, achieving dense 3D models using SGBM and filtering.

OpenCV 3D Vision Stereo Matching Calibration

• Developed a robust stereo vision pipeline using OpenCV to compute disparity maps and reconstruct 3D point clouds from stereo images, improving depth estimation accuracy.
• Optimized disparity computation by fine-tuning SGBM parameters and applying filtering techniques, reducing noise and enhancing 3D reconstruction quality.
• Generated and exported point clouds in PLY format, enabling seamless visualization and analysis in 3D modeling applications.

Github More Details

Defect Detection in Prints Using U-Net

Implemented a U-Net model for defect detection, achieving 95% precision and reducing manual inspection time by 50%.

Defect injection scikit-image Augmentation

• Implemented a U-Net-based model for defect detection, achieving 95% precision in quality assurance.
• Built an AI pipeline processing 1000+ images per day, enhancing data throughput by 25%.
• Automated defect detection reporting, delivering real-time insights to quality control teams.

Github More Details

Predictive Maintenance Using XGBoost

Built a predictive maintenance model using XGBoost, achieving high accuracy in failure prediction for industrial systems.

XGBoost Python Data filtering

• Developed an end-to-end predictive maintenance model leveraging XGBoost for industrial failure prediction.
• Achieved 98% accuracy by tuning model hyperparameters and implementing advanced analytics.
• Automated reporting and insights generation, enhancing operational efficiency.

Github More Details

Reinforcement Learning for RRR Robot

Implemented a RL model for end effector path planning, optimizing robot arm trajectories in a 3D environment.

Reinforcement Learning MATLAB Python Kinematics

• Developed and implemented RL-based trajectory optimization in MATLAB, improving end-effector accuracy by 30% in reaching target positions.
• Trained the robot over 1,000 episodes using reward-based learning, reducing path deviation by 25% and ensuring efficient motion planning.
• Simulated and validated RL policies in MATLAB, achieving an 85% success rate in autonomously selecting optimal paths for task execution.

Github More Details

ECG Signal Classification with AI

Developed a deep learning model for classifying ECG signals, achieving 99% accuracy in detecting cardiac anomalies.

ECG / Biosignals MATLAB Signal Processing

• Developed a deep learning model in Python for ECG signal classification, achieving 99% accuracy in detecting abnormal heart rhythms.
• Preprocessed and segmented raw ECG data, improving model training efficiency by 40% through feature extraction and noise reduction.
• Optimized model inference speed, reducing prediction time by 30%, enabling real-time arrhythmia detection for potential clinical applications.

Github More Details

Get in Touch

Contact Me