AI Electrical Symbol Detection on Contractor Blueprints

At a glance

The symbol detection build for a US electrical contracting business comes down to four short answers & one fact table.

Why did symbol counting stay manual?

Electrical symbol detection stayed manual at a US electrical contracting business because every bid starts with a count of devices on dense power-plan drawings.

Estimators counted receptacles & data outlets on every sheet, along with each fire alarm device. The count lived in an estimator’s eyes, because computer vision object detection had never been pointed at these sheets.

Where the hand count broke

• Each takeoff began with an estimator reading one sheet slowly, highlighter ready.
• Crowded regions hid symbols inside overlapping linework, & a missed device became a missed cost.
• Similar pairs such as duplex against fourplex symbols invited swaps that nobody caught until the material order arrived.
• Deadline pressure turned careful counting into fast counting, & fast counting is where the errors lived.
• Checking a count meant redoing the whole count, so review cost the same hours all over again.

Researchers who reviewed drawing digitisation reached the same verdict years ago, finding automatic analysis of engineering drawings still far from complete^[1].

What do estimators usually try first?

Estimators usually try one of four routes first, & each route suits some teams well.

Published research on drawing symbols reports the same obstacle, because real drawings are heavily imbalanced between symbol classes & rare devices suffer^[2].

How did we design the detection pipeline?

Brainy Neurals designed the detection pipeline around one rule. An estimator only sees a box that has earned its place.

A full drawing set goes in as one PDF, & the annotated sheets come back out. The client never has to convert or export anything first.

Our first decision was what the detectors would train on. We rejected generic symbol libraries, because legends differ from office to office even under the same electrical code. The detectors learned the client’s own legend instead, the way document AI systems learn a client’s own forms.

Our second decision was what would count as a real detection. Raw output put boxes on blank paper, & grid references kept reading as electrical devices.

Every box now passes three checks before it reaches a sheet, & overlapping finds merge into one. After that change, the ghost boxes stopped surviving.

An estimator only sees a box that has earned its place.

Mitesh Patel set the detection approach for this build & reviewed every filter that decides what an estimator finally sees.

The detectors are the least interesting part of the build, because everything around them is what made the counts trustworthy.

What technology stack runs the pipeline?

The technology stack behind the electrical symbol detection pipeline had to survive real drawing sets, including hundred-page PDFs & crowded sheets. Legends on those sheets also vary from office to office.

We chose each layer for what it removes from an estimator’s working day. Weighing routes like these is AI consulting work, done before any training run gets paid for.

Document intake

Detection

Review readiness

How does one drawing set get read?

One drawing set gets read in six fixed steps, & the estimator only arrives after the sixth.

1. The system renders each page of the uploaded PDF at print quality, so fine linework stays sharp.
2. A text scan across each page keeps the power-plan sheets & quietly sets the rest of the set aside.
3. The pipeline slices every kept sheet into overlapping tiles, so no symbol shrinks below what detectors can see.
4. Several specialist detectors scan every tile at once, each one trained on a different family of symbols.
5. Three filters strike out blank-paper marks & grid references, along with look-alike reference bubbles, before any box survives.
6. Overlapping finds merge into one, & the sheet returns with every symbol boxed & labeled, each with a score.

No step in the pipeline waits on a person, & an estimator appears only at the review after the sixth step.

What nearly stopped the build?

Three problems nearly stopped the build, & they arrived in this order.

The first detector pass found almost nothing, because squeezing a whole sheet to model size starves every symbol of pixels. Small-object research names the same failure, since tiny targets carry too little visual information for a detector to hold onto^[3].

The second pass found too much, with boxes on blank paper & grid marks like S.5 reading as devices. A count full of ghost boxes is worse than no count at all.

The third problem hid inside the training sheets themselves. Common receptacles appeared by the thousand & rare fire alarm devices by the handful, so the rare ones kept losing. On safety devices like fire alarms, that is exactly the wrong way to fail.

Clients hire AI developers for weeks like these instead of learning the same lessons at bid speed.

How we fixed each problem

We fixed each problem with a separate change, & each fix took longer to find than it takes to read.

Scale

We sliced every sheet into overlapping tiles & ran detection on the tiles instead of the page. Each symbol now reaches the detectors near full size, & a symbol cut at an edge survives in the next tile.

Ghosts

A brightness test rejects boxes on blank paper, & a text test reads each box to throw out grid references. A shape test rejects the oval reference bubbles, & the boxes that survive merge where they overlap.

Imbalance

We split symbol families across specialist detectors & rebalanced the examples until rare devices carried real weight. A missed fire alarm costs more than a false one, so safety symbols run at a deliberately sensitive setting.

The word blueprint outlived the blue printing process behind it by decades.

Surfacing problems like these before a bid depends on them is what an AI proof of concept is for.

What changed after go-live?

After go-live, the first count of every power-plan sheet moved from the estimator to the pipeline.

We publish no detector accuracy figure here. Every count still passes an estimator’s review before it reaches a bid, & that review is the standard that matters.

Day to day, an estimator now opens an annotated drawing set instead of a blank one. The counting hours became review minutes, & review is where an estimator’s judgment earns its keep.

What is running today

The electrical symbol detection pipeline runs in production today as a web application that the client’s estimators use on live bid work.

Estimators upload the PDF sets they already receive, & the annotated sheets come back ready for review.

Since handover, the pipeline has read entirely new drawing sets without any retraining. The legend it learned is still the legend the client’s construction projects use today.

The estimators’ files stayed exactly the same, & only the reading of them changed.

What would we do differently?

We would change four habits if we started this build again, & each one came from a mistake we made here.

Balance the classes before the first run

Common symbols drowned the rare ones, & we rebalanced only after the rare ones failed. The rarest symbol on a sheet is usually the one a bid can least afford to miss.

The rarest symbol on a sheet is usually the one a bid can least afford to miss.

Put the cheap filters first

A plain brightness check on blank paper removed more junk than anything clever, so the simple test now runs first.

Test on the ugliest sheet you have

Clean sheets flatter a detector, & the crowded sheet with everything overlapping is the one that tells the truth. Every build now starts on that sheet.

Let the specialist win every disagreement

When a general detector & a specialist disagree on a symbol, the specialist has been right every time. We made that a fixed rule instead of a debate.

Where else does electrical symbol detection fit?

Electrical symbol detection fits anywhere a takeoff still happens by eye, one colored pen at a time.

Automated symbol detection finds & counts the standard symbols on technical drawings by running trained detectors across every sheet.

Porting the pipeline takes annotated samples of the new legend & retrained detectors, followed by one review cycle with whoever owns the count. The filters carry over to the new drawings unchanged.

Services behind this case study

Six Brainy Neurals services sit behind this case study, from detection models to extra engineering hands.

An AI proof of concept is the fastest way to test this pattern on real sheets. AI consulting helps weigh the routes first, & an AI readiness assessment shows whether the drawings can carry it.

The AI solutions by industry page shows where this same pattern already runs today.

Similar case studies

Similar case studies from Brainy Neurals each shipped into a real working environment instead of a demo.

Cite this case study

Cite this case study with the reference below, & use the numbered sources for the research it draws on.

Patel, Mitesh. AI Electrical Symbol Detection on Contractor Blueprints. Brainy Neurals, August 2026. https://brainyneurals.com/case-studies/automated-electrical-symbol-detection/

Sources cited on this page