Capture & normalize
Acquire the PRECAL form from mobile and prepare it for reliable model inference.
Back to selected workOptiFlow
Confidential case study / Document intelligence
OptiFlow
Precal Insight
A mobile and AI system that turns photographed PRECAL optical forms into reviewed, business-ready data.
Confidential final-year engineering project delivered at Essilor SIVO (EssilorLuxottica), Sfax. Source code, product screenshots, implementation artifacts, and demonstrations are not public.
AI
95.16%mAP@50
The business problem
Manual transcription created a bottleneck between paper forms and usable data.
PRECAL optical forms contain multiple field types that need to be detected, interpreted, matched to catalogue data, and checked against business rules. The previous workflow depended on manual reading and entry.
The proposed system captures a form from mobile, routes each region to the right recognition task, pre-fills structured data, and keeps an operator in control whenever confidence or domain rules require review.
My role
End-to-end ownership across product, software, and AI
01
Requirements and architecture
02
Mobile and backend delivery
03
Dataset preparation and model evaluation
04
AI pipeline integration
05
Testing and quality assurance
06
Human-in-the-loop validation
07
Testing and quality assurance
System architecture
A staged pipeline with explicit review boundaries
Each stage has one responsibility, produces inspectable output, and can route uncertainty forward without hiding it.
Detect fields
Use YOLO11 to locate structured regions before task-specific recognition.
Route recognition
Classify empty or filled fields, read numbers with TrOCR, text with GLM-OCR, and evaluate checkboxes.
Resolve & validate
Apply RapidFuzz catalogue matching, confidence thresholds, and business validation rules.
Review & learn
Pre-fill the form, route uncertainty to a human, and retain corrections as feedback memory.
INPUTMobile capturePRECAL form
→DETECTIONYOLO11Field regions
→RECOGNITIONEfficientNetV2-STrOCR · GLM-OCRCheckbox model
→DECISIONRules + RapidFuzzConfidence gates
→OUTPUTReviewed pre-fillFeedback memory
Validation
Measured at model and workflow level
Precision
YOLO11 field detection
Recall
YOLO11 field detection
mAP@50
Detection evaluation
mAP@75
Detection evaluation
mAP@50–95
Detection evaluation
The validation program combined detector metrics with end-to-end human-in-the-loop workflow checks.
Engineering decisions
Three decisions that shaped the system
Mixed field semantics
A single form contains positional fields, empty/filled states, numerical values, free text, and checkboxes. The pipeline separates these tasks instead of forcing one model to solve all of them.
Confidence is not correctness
Model confidence is combined with catalogue resolution and business rules. Ambiguous predictions remain reviewable instead of being silently accepted.
Automation without losing control
Human-in-the-loop validation protects the final workflow and turns corrections into reusable feedback for later processing.
What I learned
Good AI engineering is system engineering.
- Decompose document intelligence into measurable, independently testable stages.
- Treat human review as an architectural capability rather than a fallback screen.
- Use domain rules and catalogue resolution to complement model predictions.
- Evaluate both model quality and the complete operator workflow.
Related writing
The reasoning behind two of these decisions
Need an engineer who can connect AI pipelines, backend architecture, and reliable product delivery?
Start a conversation