🛠️ ResoluteAI Internship Blog: How I Used Deep Learning to Solve Real Industrial Problems

When I joined ResoluteAI as a Deep Learning Engineer Intern, I didn’t just want to train models and get good accuracy. I wanted to build things that actually solved field-level problems — systems that could help real businesses save time, automate tedious work, and make smarter decisions.

Over the course of the internship, I worked on pipe inspection systems, agricultural monitoring tools, and web-based AI visualizations — applying a mix of YOLOv8, OpenCV, and PyTorch to real-world drone and camera data.

---

📍 Role: Deep Learning Engineer Intern

Company: ResoluteAI
Duration: June 2023 – April 2024
Location: Mumbai (Remote)
Tech Stack: YOLOv8, OpenCV, TensorFlow, PyTorch, Scikit-learn, Streamlit

---

🔍 Project 1: Detecting Nested Pipes Using YOLOv8 + Computer Vision

🧠 The Problem

Clients in the infrastructure industry needed a way to detect pipes nested inside one another in inspection footage. These pipes overlapped visually, had glare issues, and varied in lighting — so manual inspection was inconsistent and tiring.

⚙️ What I Built

• Trained a YOLOv8 object detection model on custom datasets to detect and classify nested pipes

• Used OpenCV techniques like thresholding, contour detection, and dilation to improve edge clarity

• Combined DL + CV logic for hybrid detection: YOLO for outer detection, CV for inner nesting

📈 Impact

• Achieved 40% increase in inspection speed

• Reduced manual review effort drastically — saving 2–3 hours per batch

• Used in a live POC demo with actual client footage

---

🌿 Project 2: Sapling Counting in Drone Imagery

🧠 The Problem

Clients in the agriculture and reforestation domain wanted to count plant saplings from aerial drone footage. The challenge? Saplings were small, often partially occluded, and had low contrast against background soil.

⚙️ What I Built

• Created a two-stage YOLOv8 model:

  • Stage 1: Detected drone frame sections

  • Stage 2: Identified individual saplings

• Trained on annotated drone frames with thousands of sapling bounding boxes

• Used Focal Loss and Non-Max Suppression to improve performance on small objects

📈 Results

• Achieved 95%+ detection accuracy

• Model was able to process hundreds of drone frames in batches

• Reduced human counting workload massively

---

🔩 Project 3: Pipe Defect Classification via OpenCV + DL

🧠 The Problem

After pipe detection, the next step was identifying defects like cracks, corrosion, or dents — even small ones.

⚙️ What I Built

• Applied thresholding, dilation, and contour-based feature extraction on grayscale images

• Built a ResNet50 classifier, trained via transfer learning on defect-type image slices

• Integrated the classifier with OpenCV's ROI bounding box workflow

📈 Results

• Reduced inspection effort by 3+ hours per defect batch

• Client appreciated its speed and integration simplicity

• Became part of the company’s POC toolkit for demos

---

🌐 Project 4: Drone Analysis Web App using Streamlit

🧠 The Problem

Even if the model works — how do you show it to a client? How do they interact with results?

⚙️ What I Built

• Developed an end-to-end Streamlit dashboard

• Let users upload drone footage → run inference → visualize detections

• Included interactive map overlays, summary stats, and download/export options

🎯 Why It Mattered

• Made the models client-presentable

• Helped managers without technical background understand what we were building

• Reduced demo prep time from 2–3 hours to 5 minutes

---

🤝 Mentoring Junior Interns

I also led and mentored 3 junior team members, introducing them to:

• YOLO model training basics

• Labeling strategies and augmentation

• Experiment logging, error analysis, and reporting

---

💬 What I Learned

• That real-world AI is rarely about accuracy alone — it’s about usability, speed, and integration

• The power of combining deep learning and traditional computer vision

• That communication is key — your model won’t make an impact if clients can’t understand or interact with it

---

✉️ Let’s Connect

If you're working on similar AI-for-industry or computer vision projects — I’d love to exchange ideas, show you what I built, or just connect!

📩 LinkedIn | 🔗 GitHub