Customer profile
The team operates within a naval architecture and marine engineering consultancy that serves chemical tanker operators, shipyards, and classification bodies across the commercial shipping sector. Their documentation work centres on DXF and DWG drawings — converting what lives in a CAD file into the structured tables, updated title blocks, and formally packaged deliverable sets that yards and classification surveyors actually need.
A typical project delivers a tank schedule, a Main Particulars summary, at minimum one title-block revision, and a return drawing in a specified DWG version. For the Hanna project — a chemical tanker — the scope covered exactly this set: two structured data tables, a title-block change, and a round-trip conversion back to a validated AC1021 DWG. The team processes multiple vessels concurrently, so the manual-extraction bottleneck compounds quickly across the project portfolio.
Problem
Extracting structured data from DXF drawings is not a text-parsing task in the ordinary sense. Ship drawings store labels, values, dimensions, and frame annotations as DXF text entities distributed across layers, with no consistent row-and-column structure that spreadsheet tools can read directly. The only reliable method is to parse the DXF format programmatically — or open the drawing in a full CAD workstation and type values out by hand.
For the Hanna project, the extraction target was two tables plus seven discrete measurement lines.
The Main Particulars block contained these vessel dimensions:
- Length for scantlings: 106.254 m
- Length overall: 109.999 m
- Breadth moulded: 13.500 m
- Depth moulded: 5.140 m
- Draft — scantling: 4.200 m
- Draft — maximum: 4.200 m
- Cargo tank capacity: 4,560 m³
Each line also carried qualifiers — "abt." for approximate, "mld." for moulded — that had to be preserved in the output alongside the numeric values and units. Losing or misassigning a qualifier changes the engineering interpretation of the dimension.
The tank table posed a different challenge. A complete tank schedule for a chemical tanker of this class spans many numbered tanks, each with a frame range and associated capacity. The risk is not only extraction error but silent omission: a missing tank entry may not be detected until the document reaches a classification surveyor or the yard's procurement team. In the team's previous workflow — opening the drawing on a CAD workstation, manually reading values, and entering them into Excel — a full extraction pass for a single vessel drawing took most of a working day and still left that risk unaddressed.
The project also required a DXF-to-DWG conversion producing a valid AC1021 file. That version standard (AutoCAD 2007 format) remains required by many older CAD systems in use at classification bodies and some yards. Simply renaming the file is not sufficient; the team needed entity counts and layout confirmation to certify the output for handover.
Why now
Classification bodies and shipyard procurement systems are increasingly requiring machine-readable data alongside drawings. A DWG or PDF alone is no longer sufficient when the receiving system needs to ingest capacity figures, frame ranges, or vessel particulars into a database or cross-check them programmatically before a survey is scheduled.
The Hanna project represented a common inflection point: a vessel transitioning from the drawing review phase to formal classification submission, where structured data exports are required in parallel with the updated drawing. Delays at this stage propagate into survey scheduling and, ultimately, the vessel's operational readiness. The team needed to compress the extraction and packaging phase without introducing errors that would require a re-submission.
Why energent.ai
The team assessed three approaches before selecting energent.ai: a dedicated CAD-workstation operator handling extraction manually, an internally maintained Python parsing script, and a general-purpose language model without direct file-handling capability.
Manual extraction remained necessary for final visual confirmation of flagged anomalies, but it was too slow and operator-dependent for the routine extraction and structuring phase. The internal Python script required ongoing maintenance as drawing structures varied between vessel types and clients. A general-purpose language model could describe DXF format conventions but could not load and parse the file itself.
energent.ai handled the complete workflow in a single session. It loaded the DXF file, ran Python to parse the text entities, structured the output into labelled columns, applied title-block changes in-place, converted the updated DXF to DWG with version validation, and packaged six deliverable files. Critically, it also ran an independent audit pass — deliberately isolated from the extraction process to avoid confirmation bias — that surfaced two anomalies before the files were delivered. That audit behaviour was the capability the team could not replicate with any of the alternatives.
Workflow
The session followed a reproducible seven-step sequence:
Step 1 — Drawing ingestion. The team uploaded the Hanna DXF file. The agent parsed the file structure and identified the three target regions: the tank schedule, the Main Particulars block, and the title block.
Step 2 — Main Particulars extraction. The agent extracted the seven-line block, matching each measurement label to its value, unit, and qualifier. Output: an XLSX workbook and a CSV, each with plain-English column annotations explaining qualifiers such as "abt." (approximate) and "mld." (moulded).
Step 3 — Tank table extraction. The full tank schedule was extracted into a structured table covering tank numbers, frame ranges, and capacity values. Output: XLSX and CSV.
Step 4 — Title-block modification. The requested title-block text replacements were applied directly to the DXF entity records. The agent confirmed the changes by inspecting the updated file's text content.
Step 5 — Independent audit. An isolated audit pass reviewed the extraction outputs against the source data. Two findings were recorded: Tank No. 12 was absent from the extracted tank table — flagged as a completeness question rather than silently accepted — and Tank 5 carried a frame range of 127–172 that appeared anomalously wide compared with adjacent entries, warranting visual review against the source drawing.
Step 6 — DXF-to-DWG conversion and validation. The updated DXF was converted to DWG using a structured conversion workflow. Validation confirmed: AC1021 version, two layouts (Model and Layout1), and 3,753 modelspace entities. A caveat was recorded that visual fidelity depends on the receiving workstation's font configuration and SHX file availability.
Step 7 — Deliverable packaging. All six files were assembled and confirmed: tank table XLSX, tank table CSV, Main Particulars XLSX, Main Particulars CSV, updated DXF, and validated DWG.
Results
Six structured deliverables were produced from a single DXF input in one session:
- Seven vessel dimensions extracted into labelled XLSX and CSV with unit qualifiers preserved
- 4,560 m³ cargo capacity correctly captured and associated with its table row
- 3,753 modelspace entities validated through the DWG round-trip (AC1021 confirmed)
- 2 audit flags surfaced before delivery — preventing silent errors from reaching downstream reviewers
- 6 deliverable files packaged, matching the standard yard and classification handover format
The team also received a reproducible session template. The same seven-step sequence — ingest, extract Main Particulars, extract tank table, modify title block, audit, convert, package — can be applied to additional vessel drawings without rebuilding the workflow from scratch for each project.
Proof
"We used to spend most of a working day on a single vessel drawing, and we still sometimes missed entries that only surfaced during the classification review. Having the agent flag a missing tank and a suspicious frame range before we packaged the files was exactly the catch we needed built into the process, not added manually at the end." — Lead Naval Architect, marine engineering consultancy
The complete deliverable set the agent produced — tank table XLSX and CSV, Main Particulars XLSX and CSV, updated DXF, and validated AC1021 DWG — mirrors the standard handover package the team submits to shipyard procurement and classification surveyors. The two audit flags indicate precisely which items require visual confirmation before sign-off, rather than leaving that determination to the downstream reviewer.
Trust note
The DWG output passed structural validation (AC1021 format, 3,753 modelspace entities confirmed), but exact visual fidelity remains dependent on the receiving workstation's font configuration and SHX file availability — a gap that cannot be resolved without a CAD workstation review. The audit explicitly flagged Tank No. 12 as absent from the extracted table and Tank 5's frame range (127–172) as potentially anomalous compared with adjacent entries. Both items require a human reviewer to open the source drawing and confirm visually before the tank table is submitted for classification review or incorporated into formal vessel documentation. The Main Particulars CSV is structurally complete for the seven core lines but does not constitute a full technical specification sheet.