From Construction Zones to Clean Labels - A Day in My Annotation Loop

A Day in the Data

Hey, I’m Faisal
At LexData Labs, I’m a Senior Data Processing Executive. On some days, the role feels more like I’m part engineer, part detective, and part perfectionist.

When I’m not in front of CVAT adjusting bounding boxes to the pixel, you’ll find me curating niche Spotify playlists (current vibe: “Chill Beats in Safety Zones”), sneaking in mini espresso breaks, and obsessing over tight label edges. Let me take you through what a typical (but never boring) day looks like.

9:00 AM - Kickstart & Task Breakdown

First thing in the morning, I log into CVAT to check for new image batches assigned to me. Today, it’s 300 raw images from an active oil rig site. From rig pits, safety cones, zone cones, A-frames, balconies to other machineries, I know exactly what I’m in for.

After a quick glance, I spot a new object we hadn’t seen before, it's a seam that’s half buried in dust. That means I’ll need to tweak the task code slightly to include it as a separate class.

Once the annotation categories are updated, I fire up my spreadsheet, this is the command center. I divide the dataset among the team, set clear targets, and include columns for remarks, confusion flags, and QA tracking.

Next, I line up my desk: dual screens ready, water bottle filled, lights dimmed just right and of course, headphones plugged in with "The Scientist" by Coldplay playing softly in the background. Yes, I’m that person who matches music mood to annotation themes.

9:30 AM - 1:00 PM Deep Label Mode

This is my quiet zone. I’m labeling machinery, workers, hazard cones all with bounding boxes so tight, not even a pixel of background peeks through. Precision is everything here.

A major challenge today? Shadows over safety cones. Sometimes they look orange, sometimes muddy brown. In one image, two cones and a zone marker had almost identical shades. After zooming in, adjusting brightness, and switching to grayscale temporarily, I finally managed to differentiate them.

I jot a remark in the sheet: 'IMG_234 – difficult to determine whether it’s a zone cone or a safety cone.' This isn’t just annotation; it’s decision-making in real time.  

Some frames take 2 minutes, others demand 15. But that satisfying moment when the box fits just right around a blurred edge? Absolute chef’s kiss."

1:00 PM - Lunch & Reset

I step away from the screen for a quiet lunch with the team. Today it’s homemade biryani, and someone brought misti doi, which we all agree should be a required post-meal tradition.

Following lunch, I take a 10-minute stroll outside. The construction noise from the nearby site feels oddly fitting, as if the world’s curating field-relevant background audio just for me.

2:00 PM - 3:30 PM Manual Quality Assurance Begins

Once my portion is complete, I dive into Quality Assurance. No scripts here it’s all eyes and experience. I double-check tightness, class accuracy, and make sure there's no overlap or missed corner.

Did I mention how unforgiving construction dust can be in image quality? In one image, a worker’s vest was nearly the same color as a rusted background beam. That took extra scrutiny. I zoomed in, compared shadows, and finally added a bounding box. The goal? No object left behind.

3:30 PM - Break & Balance

I sometimes engage in a small “box test” during breaks which involves drawing a perfect square freehand to see how aligned it is. Weirdly satisfying.

4:00 PM - 5:30 PM Team Review + Final Sweep

During this period, I check the spreadsheet. Team members mark a few confusion points. I open those images, cross-verify, and add notes.

Overall, we make sure nothing slips through. Every image we touch is part of a safety-critical system, leaving no room for ambiguity.

5:30 PM - Submission & Wrap Up

By the end of the day, our dataset is QA-passed and ready for delivery. We submit the folder along with the remark-annotated spreadsheet; Our version of a field report.

And just like that, we took raw, unstructured visual chaos and turned it into structured, AI-ready gold.

Why It Matters

🛠️ Real-time safety: Our labels help AI detect workers and hazards instantly
🚧 Faster deployment: Clients can train and iterate faster with clean data
⚙️ Zero bottlenecks: No waiting on QA our workflow moves like a well-oiled rig

We don’t just draw boxes; We define boundaries between risk and safety, between human intuition and machine vision. Every cone, every object, every edge matters.

Final Thought

I might not be on-site in a helmet, but in every labeled image, I’m building a safer, smarter oil rig. One bounding box at a time.

See you after the next 300. 🎧🖱️

‍

‍