Computer Vision for Field QA
From Manual Checks to Automated Intelligence
-
100%
Manual image review replaced by automated computer vision
-
230 GB
Inspection data processed — images, measurement files, and protocols
-
13+
Independently trained object-detection models in the library
Executive Summary
Primotly designed and built a modular AI inspection platform for a German management and technology consulting firm specialising in broadband infrastructure. The platform replaces a fully manual installation quality assurance process — where every photograph, measurement file, and protocol document was reviewed by a human analyst — with an automated computer vision pipeline that detects, classifies, and validates fiber network assets in minutes, at any scale. At the core of the platform is a library of 13+ independently trained object detection models, each responsible for one component type — distribution boxes, optical terminals, splice cassettes, cable entry points, and more. Models compose into job-specific inspection pipelines on demand, configurable without retraining. A smart pre-classifier routes each image to only the relevant models, and secondary compliance-rule checks verify installation correctness beyond mere detection. A single address directory that previously required 2–3 minutes of manual analyst review now processes in under 10 seconds — more than 12× faster. A standard job covering 100+ directories, which once required 3–5 hours of continuous human review, now completes in under 20 minutes. In one city deployment, the system surfaced a systemic compliance gap — missing identification labels on approximately 90% of inspected devices — that had not been caught at scale under manual review.
About the Client
The client is a German management and technology consulting firm specialising in broadband infrastructure and fiber network deployment. They serve regional and local telecommunications providers across two core divisions: a consulting arm covering broadband strategy, operational excellence, and M&A advisory; and an engineering arm covering network planning, construction supervision, and end-to-end project management. Beyond advisory and engineering services, the client develops proprietary digital products — including AI-powered document verification tools and workflow management platforms for fiber infrastructure — positioning themselves at the intersection of consulting expertise and technology-driven service delivery. At the time of engagement, their installation QA process was entirely human-driven, and the business was growing faster than the workflow could scale.
- AI/ML Development
- Custom Software Development
Telecommunications / Broadband Infrastructure
Fixed Price — Extended Engagement
7 people at peak
The Challenge: A Workflow That Could Not Scale
Every installation job produced hundreds of photographs, measurement files, and protocol documents — all reviewed by hand. As the business grew, the manual QA workflow hit a rigid ceiling across three fronts:
- 01/
Growth Constrained by Headcount
Analysts had to visually inspect each image to confirm whether the correct equipment was installed, properly connected, and compliant with specifications. The process was slow, expensive, and impossible to scale: doubling inspection jobs meant doubling analyst headcount. Without automation, growth was directly constrained by the number of people available to review.
- 02/
Inconsistency and Compliance Risk
Human fatigue introduced inconsistency, and there was no systematic way to catch all defect types reliably across large datasets. Missed defects from reviewer fatigue or inconsistent interpretation could result in failed audits and damaged client relationships — a direct compliance risk sitting on top of the process.
- 03/
Sequential, Human-Bottlenecked Turnaround
The inspection workflow was entirely sequential: images collected in the field, uploaded, downloaded by analysts, reviewed individually, results entered manually. Turnaround time from data collection to compliance report was measured in days. The client recognised the limitation but lacked the internal capability to build a computer vision solution independently.
The Primotly Solution: Composable AI Inspection
Primotly designed and built a modular AI platform in which each physical asset type is detected by a dedicated, independently trained computer vision model. These models compose into job-specific inspection pipelines on demand — configurable without retraining, without architectural changes, and without per-job custom development.
The guiding principle: automate every step a human analyst previously performed manually, with confidence-scored outputs and a defined escalation path for edge cases.
The engagement began with a structured discovery phase — requirements workshops with client stakeholders, field data analysis, feasibility assessments for high-risk detection scenarios, and modular architecture design sessions. A full capability map was produced: every object type that previously required manual identification was catalogued, assessed for technical risk and data availability, and prioritised before any implementation began.
Significant R&D effort was absorbed internally by Primotly and not billed to the client, demonstrating genuine technical partnership. All high-risk detection scenarios underwent formal feasibility assessment before scope commitment, preventing over-promising and protecting delivery credibility throughout the engagement.
Key Features
-
Modular AI Pipeline Configurator
The required detection capabilities are selected from a library of models, and a ready-to-run automated inspection pipeline is configured and deployed — replacing what was previously a manual analyst workflow. New inspection configurations can be deployed in days rather than requiring weeks of setup, letting the business take on more jobs and serve more clients without proportionally expanding the analyst team.
-
Multi-Component Object Detection Engine
A library of 13+ specialised computer vision models, each trained to detect a specific physical asset: distribution boxes (open, closed, covered), optical terminals, subscriber connection boxes, splice cassettes, cable entry points, microducts, cable protection ducts, and more. Each model operates independently, is versioned separately, and composes into inspection pipelines as required — processing hundreds of images in the time it would take an analyst to review a handful.
-
Automated Document & OTDR Compliance Analysis
Automatically classifies protocol PDFs, extracts structured data fields, and processes OTDR measurement files (binary SOR format) to validate optical signal levels against compliance thresholds. Results are aggregated into a standardised, auditable compliance report delivered alongside image analysis outputs — replacing hours of manual document review per inspection job.
-
Smart Pre-Classifier & Image Routing
A lightweight AI classifier that analyses each uploaded image and routes it exclusively to the relevant detection models — rather than running the full model library on every image. Operating at an 85%+ confidence threshold before falling back to full pipeline execution, it significantly reduces processing time and cloud infrastructure cost for large-scale, high-volume inspection runs.
-
Automated Compliance Rule Verification
Goes beyond detection to verify whether detected objects meet installation rules. Once assets are identified, the platform applies a configurable set of compliance checks: is a specific cable physically connected to the correct device? Are pigtails installed in the correct positions? Is the required component present in the expected location? Rules are evaluated automatically on every processed image, with violations flagged in the output — the difference between knowing what is in an image and knowing whether the installation is correct.
Technical Deep Dive
- 01/
Separating Visually Indistinguishable Devices
-
The Problem: Two distinct fiber installation components — a Gf-TA subscriber connection box and an ONT optical terminal — are visually indistinguishable to standard computer vision classifiers. Both devices can appear in the same installation image and require separate, correct identification for compliance reporting. A misclassification has direct regulatory consequences.
-
The Solution: Both models were trained with explicit negative class examples of the confusable device, and training data was pooled across multiple client engagements to maximise variant coverage. Confidence thresholds route low-certainty detections to a human review queue rather than producing a silent incorrect classification, and a formal feasibility assessment was conducted before committing the detection task to scope.
-
The Result: reliable device separation in production conditions, with a defined human escalation path for edge cases and transparent detection confidence in every output record.
-
- 02/
Training Reliable Models on Extremely Limited Data
-
The Problem: Several device types had extremely limited real-world training images — in some cases as few as 30 examples per device type, against a minimum requirement of 200–1,000 labelled images for reliable model training. Standard training approaches would have produced unreliable models and misrepresented confidence to the client.
-
The Solution: Internal data augmentation pipelines were built to expand limited datasets, and an auto-labelling tool was developed to accelerate dataset creation from raw field imagery. Synthetic data generation was used for device types where real examples were unavailable, and explicit feasibility studies were conducted before accepting detection tasks into scope — with all estimates structured around clear data-dependency conditions.
-
The Result: models deployed with clearly defined accuracy targets and data prerequisites, and a data-conditioned pricing model that lets the client understand cost drivers and supply data proactively.
-
- 03/
Architecting a Composable 13+ Model Platform
-
The Problem: Building a computer vision system to replace manual inspection is not just a model training problem — it is an architecture problem. The solution needed to support 13+ independent detection models, be configurable per job without retraining, and remain maintainable as new object types and device variants were added over time. No established template existed for this domain.
-
The Solution: A single-responsibility principle was defined for all models: one model per object concept, trained on pooled field data, with standardised input/output contracts. A pipeline composition layer treats models as interchangeable components, while a pre-classifier routing layer orchestrates model execution efficiently — an approach validated by 40+ hours of internal R&D before implementation.
-
The Result: a fully composable platform where new inspection configurations deploy in days, model updates roll out without pipeline reconstruction, and the detection library grows incrementally. The architecture is domain-agnostic and directly transferable to any field inspection use case.
-
Business Impact & Results
The composable inspection platform delivered measurable gains across processing speed, defect detection, scale, and the unit economics of quality assurance.
-
12× Faster Directory Processing
A single address directory that required 2–3 minutes of manual analyst review now processes in under 10 seconds — more than 12× faster. A standard inspection job covering 100+ directories, previously 3–5 analyst hours, now completes in under 20 minutes: an over 90% reduction.
-
Zero Manual Data Entry
No manual data entry is required for asset identification, label reading, GPS traceability, or measurement validation — all extracted automatically. Binary measurement file processing is fully automated: SOR format parsed, values validated, non-compliance flagged with no analyst involvement.
-
Systemic Defects Surfaced at Scale
In one city, the system identified a missing identification label on approximately 90% of inspected devices — a systemic compliance gap that had not surfaced at scale under manual review. It also flagged an unfamiliar cable connection configuration across multiple jobs, confirming a previously undocumented installation variant before it propagated further.
-
100% of Manual Review Replaced
100% of manual image review was replaced by automated computer vision across all active inspection jobs, with 13+ object detection capabilities covering the full range of previously manual identification tasks and over 230 GB of inspection data processed to date.
-
Changed Unit Economics of QA
The platform architecture scales to 10× current job volume without structural rework, and growth in inspection capacity no longer requires proportional analyst headcount — fundamentally changing the unit economics of quality assurance. The reusable component library is transferable to manufacturing QC, food production defect detection, construction compliance, and infrastructure asset management.
Tech Stack
- Frontend
- React (web application)
- Backend / ML
- Python (ML inference, pipeline orchestration)
- Infrastructure
- AWS
- Key Libraries
- Computer vision / object detection frameworks, OCR, OTDR binary SOR parsing, PDF extraction, reverse geocoding API
- Process & PM
- Scrum, ClickUp, Slack, Microsoft Teams
/ Case studies /
See our case studies
How can we help you?
Let's find the best solution for your business together
Book a meetingHow can we help you?
Let's find the best solution for your business together