The case for deterministic vision

Reliable inspection — not probabilistic guesses

Vision-LLMs are probabilistic and billed per call: same photo, different answer, every time. Deterministic computer vision is repeatable, audit-friendly and runs on free open-source weights — zero per-call cost, predictable accuracy, full data control on your own infrastructure.

Accuracy above threshold

100%

0 errors at ≥90% confidence

Per-photo inference

~10s

Single-frame ML CV pipeline

Our stack

Cost per 1 000 photos

€0

Self-hosted ML CV — no per-call API fees

LLM API

Cloud vision API

€15+

Same volume via cloud vision LLM

Prototype stack
  • Ultralytics YOLO
  • PaddleOCR
  • FastAPI
  • Nuxt 3
  • Folium
  • SQLite

Pipeline output

Four inspection categories the model reports back

Each frame is graded by what's actually in it — cable (duct) and measurement reference (tape), both, one, or neither. The category tells the inspector whether the report stands on its own or needs a re-shoot.

1Full context
ducttape

Cable AND measurement reference visible — the report stands on its own.

2Cable only
ducttape

Cable visible, measurement reference missing — partial context.

3Measurement only
ducttape

Measurement in frame, cable not visible — partial context.

4Nothing in frame
ducttape

Neither cable nor measurement — flag for re-shoot.

01

Take Original Photo

The process starts with the unmodified field image captured during inspection.

02

Apply YOLO Detection

YOLO identifies relevant visible objects, pipe elements, and structural features in the photo.

03

Categorization

Detected objects are grouped into meaningful categories based on the recognition results.

04

Apply OCR For Location

OCR reads embedded photo information to retrieve latitude and longitude when available.

05

Evaluate Confidence Level

The system scores recognition certainty so clients can understand the reliability of the output.

06

Generate Report

Final findings are assembled into a structured report for review, export, and follow-up decisions.

Detection in action

Real photos, real confidence — scales with the labelling budget

Boxes left, distribution right. Drag the slider to see how the confidence curve sharpens as the training pool grows.

Original field photo

Original Photo

Field photo after YOLO detection

After YOLO Detection

Confidence histogram

YOLO confidence — correct vs false

Threshold ≥ 90% — no errors above this confidence. Sample size n=394 (classes 1-3).

56
90-100
100
80-90
10
61
70-80
10
43
60-70
2
38
50-60
4
21
40-50
1
33
30-40
3
10
20-30
2
10-20
0-10
Training photos200
140 train·40 val·20 test·70 / 20 / 10 split
Correct (real == YOLO) False (real ≠ YOLO)

Hands-on demo

See the pipeline run on your own photo

Try it live — watch the walkthrough, then upload an inspection frame and get a structured report back in seconds. Code & Documentation on GitHub.

Open the prototype