This was my TN10 final internship at ART X BAT, working in MOEX / OPC — execution-phase client representation, scheduling, coordination and site supervision.
I followed three very different operations: a new build, an urban renovation, and the refurbishment of office floors in a high-rise (IGH) building at Paris La Défense (floors 34–39, Tours A/B). The high-rise became the main case study.
The site constraints were concrete:
strict access, visitor and authorisation control;
limited delivery dock, delivery slots and storage;
heavy trade co-activity and many interfaces;
highly repetitive office layouts with regular ceiling grids and technical equipment;
a tight schedule and a surge of punch-list items during OPR (pre-handover operations).
In this context, MOEX/OPC is not about “recording progress”: it is about turning site information into actionable work — locate, describe, assign, follow up, verify.
What I did
Site tracking
Attended meetings, took notes and clarified technical points;
checked headcount by trade against the planned progress;
carried out site inspections, photos and measurements;
tracked badges, access rights and technical-area authorisation;
coordinated the delivery dock, transport slots, storage, waste and site traffic;
updated progress tables by floor and zone;
wrote clarification, reminder and traceability emails.
Quality and punch list
Identified and consolidated punch-list items by trade;
structured them in IDCapture: location, package, contractor, description, deadline, photo;
used filters and statistics to prepare follow-up lists;
joined OPR inspections and published consolidated lists;
set up a dispute form to create a traceable client/contractor process.
Tool rollout
mapped participants and subcontractors into an account list;
checked users, roles and permissions;
prepared cleaned, zone-split PDF drawings;
designed a three-level contractor / MOE / client validation workflow;
configured the “site tracking / OPR / handover” sessions;
coordinated with the software vendor;
introduced the tool to contractors, shared tutorials and organised training.
The problem: a punch-list item without a location fails
A useful item answers five questions:
Where: floor, zone, drawing, reference point;
What: a sufficiently technical description;
Who: the responsible package and contractor;
When: deadline or target date;
Evidence: photos, status history, acceptance record.
During OPR the number of items rises quickly. In repetitive office spaces, the same defect can appear dozens of times. To reduce input, teams often create a “general item” (réserve générale) grouping several occurrences into one record.
On paper this is efficient. On site, the record loses its precise location: a contractor may fix some occurrences and miss others, while the MOE cannot check how many points existed or whether all of them are closed.
The result: input becomes easier, but verification, reminders and communication become harder; the risk of omissions, duplicates and untreated items rises.
OpenCV prototype
The goal was not to replace people, but to test whether a site photo can produce usable location information. I built two pipelines.
Case A — missing sprinkler rosette
The test image contains a manual red circle:
convert BGR to HSV;
isolate the two red hue ranges around 0° and 180°;
merge the masks;
apply closing and opening;
extract external contours and filter small areas;
convert the main contour into a bounding box.
Case A: the red annotation becomes a candidate bounding box. This validates segmentation → contour → location; it does not detect the defect itself.
The chain is stable when the visual signal is strong, but it detects the annotation, not the missing rosette.
Case B — open ceiling plenum
This case is closer to reality and does not rely on annotation:
grayscale and Gaussian blur;
blackhat with a wide rectangular kernel to enhance dark joints;
Canny edges;
HoughLinesP for horizontal and vertical segments;
merge nearby positions and rebuild the ceiling grid;
split into cells and compute mean, standard deviation and Laplacian texture energy;
select “abnormal” cells;
inside those cells, use blackhat, Otsu thresholding, morphology and area/aspect-ratio/local-contrast filters to find elongated dark openings.
Case B: candidate openings on the ceiling grid. Results are fragmented; shadows, ducts and lighting changes create false positives.
The pipeline produces candidate regions, but one opening may produce several boxes. Site visuals vary too much for fixed rules; a punch-list item is a contractual and engineering concept, not just a dark area.
Results and limits
Assisted localisation is useful: a photo can become a mask, contour, candidate box or grid cell — the first step towards pre-filling an item.
A single detection is not enough: a representative annotated dataset, a defect taxonomy by package, an independent validation set and acceptance metrics are required.
The reasonable automation is assistance: propose a zone, pre-fill a record, then let the MOE confirm.
Supervised learning was not trained during the internship because of time and missing annotations; the report frames it as a perspective, not a delivered result.
Public poster: problem, method, experiment and limits. Original deliverables have been redacted.