Built for the people who actually have to produce the data
CCTV CONTRACTORS
Drowning in video backlog
Inspect more feet per day without hiring coders or buying software your crews have to learn.
CONSULTING ENGINEERS
Data you can defend
PACP-coded, QA'd, and documented ready for a rehabilitation prioritization your client will sign off on.
MUNICIPALITIES
Use the platform you already own
Your GIS and asset system stay the system of record. PipeAid just makes sure the data going in is right.
Where we stand today:
Defects
88
F1 Score
176
What's an F1 score, anyway?
An F1 score measures two things at once: how often the model is right when it flags something (precision), and how much it actually catches out of everything it should have found (recall). F1 blends the two into a single number between 0 and 100.

Why it matters.
A model can look accurate for the wrong reasons. One that only flags defects it's very confident about will rack up a high "I was right" percentage, but it's quietly missing real problems in the pipe.
How it's measured.
We run the model against inspections it has never seen or trained on. A detection only counts as correct when the model gets the defect code right and places it at roughly the right spot in the pipe.
What counts as a good number.
Scores in the 90s are excellent, the 80s are strong production-ready performance, and lower starts to mean the model is missing things. Our current 88 across 176 defect codes puts us solidly in that strong range.
Want the full technical breakdown? Google's plain-language explainer is a solid primer.
Why "AI coding" alone isn't enough
"97% accurate" means nothing without a denominator. An F1 score tells you how often a model both finds a defect and doesn't invent one. Most vendors won't show you theirs.
We publish ours, we show the model's reasoning on every call, and a PipeAid coder checks the output before it ships. Escalations go to a PACP-certified reviewer.
Ask any AI coding vendor for their F1 by defect code. If they won't share it, that's your answer.
PUBLISHED ACCURACY
88 F1
across 176 PACP defect codes
EXAMPLE: ONE CODED DEFECT WITH REASONING

IRJ — Infiltration Runner Joint · 37.06 ft · 10 o'clock to 1 o'clock
Model: visible infiltration at the top of the pipe with additional staining from 7 o'clock to 5 o'clock. no visible structural defects. → IRJ not IGJ. Confidence 0.91.
Verified by PipeAid coder
The honest comparison
Inspection platforms are good at managing inspections. They were not built to guarantee the coding inside them. That's the part we do.
The question you're asking
Inspection platform + AI coding
PipeAid
A software subscription, plus an AI add-on
AI only, you QA it humans optional
Globally
New platform = retraining staff, installing, troubleshooting and integrating
Seats and annual license
Weeks of setup and training
Rarely published, hidden behind inflated numbers of training footage
Inside yet another platform
Replace it
A finished data deliverable
AI + PipeAid's NASSCO technicians verify every run
Domestically in Columbus, Ohio
Nothing. Send video, receive data.
Per linear foot aka pay for what you inspect
First batch back in less than 1 week
F1 score published, reasoning shown for every defect call
PACP database + GIS-ready layers, yours to keep
Keep it
What am I buying?
Who does the coding?
Where are the humans located?
What changes in my workflow?
How is it priced?
How long to get started?
Do you have model statistics?
Where does the data live?
My existing software?
How it works
1
Send us video
Upload, share a cloud folder, or ship drives. Any camera, any platform export.
2
AI codes, humans verify
Our vision-language model codes to NASSCO PACP. PipeAid coders QA every run and resolve the hard calls.
3
Data you can use
PACP database plus GIS-ready layers. Import to your platform of choice or straight into ArcGIS.
Turnaround: 1 week Pricing: per linear foot · no subscription
What to look for when choosing a provider
Questions worth asking any AI sewer inspection provider, including us.
✓
Will they publish an accuracy number?
If not, assume it’s because the number wouldn’t help their case.
✓
Full PACP standard, or a shortlist?
“AI-powered” often means “AI-powered for the easy stuff.”
✓
Who reviews the model’s work, and where?
Human oversight matters. So does its accountability.
✓
How was it evaluated, and by whom?
Self-reported accuracy isn’t the same as independent scoring.
✓
Volume vs. precision?
“Miles trained on” is a vanity metric.
✓
Was it built for sewers?
General-purpose AI can describe a pipe. It wasn’t trained to code one.
"But we already have a platform."
Good. Keep it. Here's how PipeAid fits next to it.
We already use ITpipes / WinCan / PipeLogix.
Keep using it. PipeAid delivers in the formats your platform imports.
We're already paying for AI coding inside our platform.
Use PipeAid as the QA layer on it or run a head-to-head on 5,000 feet and compare the two blind.
Our contractor does the coding.
Then PipeAid lets your contractor inspect more feet per day.
Is my data locked in?
No. You own the deliverable, in standard PACP and GIS formats.
Does this meet our NASSCO / spec requirements?
Yes. We can document our QA process and coder credentials for your spec.
CASE STUDY
Three Operators, Zero Disruption
One municipality went through three operators in a single season, from a decade-long veteran to a first-week hire. PipeAid's coding stayed accurate and consistent the whole way through.
“
The biggest surprise was what didn’t change. Same software, same crew, same workflow. What did change was the data: it’s consistent across every inspector, and I can actually trust it for capital planning.
Public Works Director, Ohio municipality
See it on your own pipes.
The best way to evaluate an AI model isn't a spec sheet. It's your own inspection video. Send us 5,000 feet you've already coded. We'll code it blind, and you compare the two. No platform to install, no contract to sign.
• Any footage, any platform export
• Results back in 1 week
• You keep the data either way
