AI that reads the drawings other AI can't.
Construction drawing sets, wiring diagrams, schematics, P&IDs and yard photos. We turn them into detections, indexed knowledge and labeled data, with every result traceable to the sheet it came from.
A vision layer that finds every symbol on a drawing, added to a quoting system that already worked.
RenderDraw turns construction drawing sets into priced line items. Iron Rod Systems is extending that platform into the electrical industry, and built the RF-DETR vision and image detection layer into it, so counts are cross-checked against what is actually drawn.
An RF-DETR detector trained per drawing set on auto-generated labels, no hand labeling. Sliced inference so small symbols on large sheets are not missed.
When no model exists yet, symbols are matched against the set's own legend, so a first pass produces usable counts before any training.
Corpus upload, legend parsing, template-matched proposals, in-browser labeling, tiled datasets, training on rented GPUs (vast.ai) or your own hardware, and per-class precision, recall and calibration.
Each symbol in the library carries an A to F grade, so the quoting team knows which detections to trust before they bid.
Corrections made during review feed straight back into the next training run.
Every box links to the sheet and location it came from, so a count can be audited against the drawing.
A 19-sheet MEP set for a Manhattan office floor, run through RenderDraw's quoting pipeline with the new detection layer cross-checking the counts.
98% accuracy: 104 of 106 symbols found on drawing regions the detector had never seen, after 74 minutes of training on a rented vast.ai GPU with auto-generated labels and no hand labeling.
Every detection traced to its location on the drawing, and every line of the resulting bill of materials to its source.
Schematics and technical documents into a knowledge base, with zero data loss.
The lineage behind Elle's patent-pending visual intelligence: vision models that classify and interpret technical pages before anything is chunked or indexed.
Schematic, wiring diagram, blueprint, civil drawing or text, decided by a vision model before OCR, so each page gets the right treatment.
Automatic retry, checkpoint resumption and integrity verification. Partial ingest states cannot exist.
Deployed as containers in the client's own AWS with private networking. Sinks for OpenSearch, Azure AI Search and Pinecone. Sources from S3, SharePoint and APIs.
Progress, failures and completions reported to Slack. Versioned releases with checksums.
Find and label equipment in thousands of photos, then export verified training data.
Built for utilities: drone inspection photos of power poles and lines, and substation equipment in yard photos. Applies to any asset-heavy operation with a camera or a drone.
Drop up to 1,000 images. A vision model finds equipment, draws bounding boxes and records type, subcategory, confidence, condition and any legible text.
A review gallery to correct boxes, reassign types and mark items verified. Zoom-aware editing.
Poles, crossarms, insulators, conductors, circuit breakers, transformers, switchgear, arresters, CTs, VTs, busbars, structures. Add your own with categories and manufacturers.
Gemini, GPT, Claude or open-weight vision models, selectable per batch from the admin panel.
Visible text defaults to none unless it is absolutely clear in the image. The tagger never fills in what the camera didn't capture.
Verified boxes and labels exported as JSON or CSV, ready as training data for detectors or as asset inventory.
- Input
- Up to 1,000 images per batch, 5 GB upload
- Throughput
- About 5 to 10 minutes per 100 images
- Interfaces
- Web app and CLI for scripted runs
- Deployment
- Containers on any cloud
- Status
- Deployed pilot, available for engagements
The rule that makes it trustworthy
AI models name, explain and draft. They never count. Every quantity comes from the drawing file and a stated rule. Every result is traceable.