Engineering Diagram Graph Annotation

About OpenTrain

OpenTrain is the #1 platform for finding and building careers in AI training and data labeling. We connect contributors to projects where human work directly shapes how AI models learn and behave.

About AI training and why this work matters

AI training (data labeling / annotation) is the human side of building modern AI: people prepare and review examples that models learn from. This project places you at the cutting edge of that process, doing remote, flexible contract work that directly improves graph-based engineering models.

• 100% remote, flexible contract work suitable for part-time contributors.
• Accessible to contributors who can read technical diagrams and follow a written guideline.

The role

You will annotate engineering diagram images as graphs: draw tight bounding boxes around each symbol (node), classify each node, and label edges that represent connections between nodes.

We will start with a small paid pilot task to confirm quality, then proceed with the full batch (~60 sheets). Final export format will be decided later; Label Studio JSON, CVAT XML, or COCO are expected targets.

• Node classes (assign one of seven): valve, pump, instrumentation, general, tank, arrow, inlet/outlet.
• Routing/helper nodes: connector (line bends/junctions) and crossing (line crossings).
• Edges between nodes labeled as solid or non-solid.
• Boxes must be drawn at full resolution and aligned to the supplied images.

What you'll do day-to-day

Perform dense, pixel-accurate annotation on large engineering-sheet images following the project guideline. Focus on consistent class assignment and exact box placement, plus accurate relationship/edge labeling.

• Draw tight bounding boxes around each symbol glyph, avoiding inclusion of connecting lines.
• Create connector/crossing helper nodes where lines bend or cross as specified by the guideline.
• Link nodes with edges that indicate connectivity and set each edge attribute to solid or non-solid.
• Export annotations in the required format and coordinate frame; label names and structure must match the provided guideline.

Requirements

You must meet the following must-haves to be considered for the pilot and full batch. Be prepared to describe concrete past experience during screening.

• Comfortable reading technical/engineering diagrams and reasoning about symbols vs. connecting lines.
• Hands-on experience with an annotation tool that supports bounding boxes and region-to-region relations (examples: Label Studio, CVAT).
• Consistent, detail-oriented labeling habits and the discipline to follow a written taxonomy and guideline.
• Ability to place pixel-accurate, tight bounding boxes on large images and review at 100% resolution.
• Experience labeling relationships or graph structure (edges), not just objects; understand edge attributes like solid vs non-solid.
• Nice to have: prior experience with object-detection or graph/relationship datasets.

Screening questions, red flags & pass criteria

We will assess tooling proficiency, diagram comprehension, classification consistency, bounding-box precision, connectivity labeling, throughput/reliability, and responsiveness to QC feedback. Strong, specific answers and examples will score higher.

• Tooling proficiency: Which annotation tools have you used, and on which ones have you drawn both bounding boxes and connections/relations between regions? (Strong answer: names Label Studio, CVAT, or equivalent and explains configuring labels and linking regions.)
• Diagram comprehension: When two symbols are joined by a line, how do you decide where each symbol's box ends and the connecting edge begins? (Strong answer: box encloses the symbol glyph only; the edge represents the link.)
• Classification consistency: How do you handle a symbol that doesn't clearly fit any class? (Strong answer: flag for review / request clarification and document ambiguous cases.)
• Bounding-box precision: What's your approach to keeping boxes pixel-accurate on large, dense images? (Strong answer: zoom to 100%, place edges tightly, use consistent margins, and self-check.)
• Connectivity labeling: Describe a time you labeled relationships or connections between objects, not just objects themselves. (Strong answer: concrete example of relation/edge or graph labeling, with awareness of edge attributes.)
• Throughput & reliability: How many densely-labeled sheets can you complete per week at high quality, and how do you keep quality consistent? (Strong answer: realistic estimate plus self-QC, rest breaks, and batch checks.)
• Feedback & QC responsiveness: If QC flags issues, how do you respond? (Strong answer: accept corrections, update annotations, and implement systematic fixes.)
• Auto-screen / red flags: No relation/connection experience, cannot name a supported tool, guesses on ambiguous symbols instead of flagging, unrealistic throughput claims without QC processes.

Deliverables, format, and payment

Deliver clean, consistent labels matching the written guideline. Exported files must be in the same image coordinate frame as the supplied images. The final tool and export format will be decided with the contractor; expected formats include Label Studio JSON, CVAT XML, or COCO.

This is a contract role paid as a fixed-price engagement ($499). We will begin with a paid 1-sheet pilot to confirm quality before authorizing the remainder of the batch (~60 sheets).

• Provide annotations aligned to the original image coordinates; we will handle downstream conversions if needed.
• Contractors are responsible for following the guideline, consistent naming, and timely responses to QC.
• Start: paid pilot (1 sheet) → approve → continue to full batch.