AI Orchestrator

A Specialized and Secure AI Orchestrator for Swiss Financial Compliance

View the Project on GitHub Digital-AI-Finance/wecan-innosuisse-ai-draft

WP5: Intelligent Document Pre-Filling

Overview

Attribute	Value
Duration	M14-M24
FHGR Hours	600h
Wecan Hours	800h
Total Hours	1,400h
Lead	Wecan Tech Lead

Objectives

Develop complete document pre-filling system for PDF and Excel
Handle variable document structures (text, checkboxes, tables, dropdowns)
Achieve >90% accuracy on 300+ forms
Deploy at 3-5 Swiss financial institutions
Create open benchmark dataset for reproducibility

Technical Approach

Pre-Filling Pipeline

Mapped Fields            Document Template
     |                          |
     v                          v
+----------+              +----------+
| Field    |              | Template |
| Values   |              | Analysis |
+----------+              +----------+
     |                          |
     +------------+-------------+
                  |
                  v
         +----------------+
         | Field Location |
         | Detection      |
         +----------------+
                  |
                  v
         +----------------+
         | Value          |
         | Insertion      |
         +----------------+
                  |
                  v
         +----------------+
         | Validation     |
         | & Formatting   |
         +----------------+
                  |
                  v
           Pre-Filled Document

Supported Field Types

Type	Detection Method	Filling Method
Text fields	PDF form fields, bounding boxes	Direct insertion
Checkboxes	Visual detection	State toggle
Dropdowns	Form field enumeration	Value selection
Tables	Structure detection	Cell-by-cell
Date fields	Format detection	Normalized insertion
Signatures	Placeholder detection	Image placement

Activities

M14-M16: Foundation

Activity	Owner	Output
Design pre-filling architecture	Wecan	Architecture doc
Implement PDF form detection	Wecan	PDF module
Implement Excel cell detection	Wecan	Excel module
Integrate with WP4 mappings	Both	Integration layer
Build review interface	Wecan	Review UI

M17-M20: Development

Activity	Owner	Output
Handle variable structures	Wecan	Adaptive filling
Implement validation rules	Wecan	Validation engine
Build confidence display	Wecan	UI component
Validate on 150 forms	FHGR	Validation results
Deliver complete system	Wecan	D5.1

M21-M24: Validation & Release

Activity	Owner	Output
Create open benchmark dataset	FHGR	D5.2
Validate on 300+ forms	Both	D5.3
Deploy at pilot institutions	Wecan	Production systems
Document results	FHGR	Final reports

Deliverables

ID	Deliverable	Due	Owner	Status
D5.1	Complete document pre-filling system	M20	Wecan	Complete
D5.2	Open benchmark dataset	M22	FHGR	Complete
D5.3	Accuracy validation (300+ forms)	M24	FHGR	Complete

All deliverable templates complete. See deliverables/ for detailed templates.

Document Format Support

PDF Documents

Feature	Support Level	Notes
AcroForm fields	Full	Native form filling
XFA forms	Partial	Legacy format
Flat PDFs	Full	OCR field detection
Scanned PDFs	Full	Visual field detection
Multi-page	Full	Page navigation

Excel Documents

Feature	Support Level	Notes
Cell values	Full	Direct cell access
Named ranges	Full	Semantic mapping
Data validation	Full	Dropdown lists
Formulas	Preserve	Don’t overwrite
Protected sheets	Partial	Unprotect if possible

Word Documents

Feature	Support Level	Notes
Content controls	Full	Native form fields
Form fields	Full	Legacy fields
Tables	Full	Cell-by-cell
Plain text	Full	Placeholder replacement

Review Interface

User Workflow

1. Upload Source Document(s)
          |
          v
2. View Extracted Fields
   [Field 1: Value] [Confidence: 95%]
   [Field 2: Value] [Confidence: 72%] <-- Highlighted
          |
          v
3. Review & Correct
   - Edit low-confidence values
   - Confirm high-confidence values
          |
          v
4. Select Target Template
          |
          v
5. Preview Pre-Filled Document
          |
          v
6. Export / Save to CRM

Confidence Display

Confidence	Display	Action
>90%	Green	Auto-accept
70-90%	Yellow	Review suggested
<70%	Red	Manual entry required

Objective Alignment

Objective	WP5 Contribution
OBJ1: 90% Document Accuracy	Final system accuracy
OBJ4: < 2 Hours Processing	End-to-end time
OBJ5: 3-5 Institution Deployments	Primary owner
OBJ6: TRL 7	Production deployment

GitHub Issues:

Benchmark Dataset (D5.2)

Dataset Composition

Category	Documents	Fields	Notes
KYC Forms	50	~2,500	Anonymized
Regulatory Filings	40	~2,000	Anonymized
Compliance Questionnaires	30	~1,500	Anonymized
Annual Reports	30	~1,000	Anonymized
Total	150	~7,000

Anonymization Process

Remove personal identifiable information (PII)
Replace company names with synthetic names
Generalize dates (year only)
Randomize numeric values within realistic ranges
Preserve document structure and field relationships

Release Format

Documents: PDF, Excel, Word formats
Ground truth: JSON annotations
Schemas: JSON schema definitions
Scripts: Evaluation code (Python)
License: Creative Commons BY-NC-SA

Pilot Deployment

Institution Pipeline

Phase	Institutions	Duration	Activities
Alpha	1 (Wecan internal)	M16-M18	Internal testing
Beta	2 (design partners)	M18-M20	Limited production
Production	3-5 (customers)	M20-M24	Full deployment

Onboarding Process

Schema Integration (Week 1-2)
- Collect CRM schema
- Configure field mappings
- Test with sample documents
System Setup (Week 2-3)
- Deploy on-premise or connect to hosted
- Configure authentication
- Set up monitoring
Training (Week 3-4)
- User training sessions
- Admin configuration
- Documentation handover
Go-Live (Week 4+)
- Supervised production use
- Feedback collection
- Iterative improvement

Success Metrics per Institution

Metric	Target
Documents processed	50+
Accuracy	>90%
Time savings	>80%
User satisfaction	>4/5

Milestone Checkpoints

MS4 (M16)

Pre-filling architecture designed
PDF and Excel modules implemented
Integration with WP4 complete

MS5 (M20)

D5.1 Complete system delivered
3+ pilot institutions onboarded
Security audit passed
OBJ5, OBJ6 validated

Project End (M24)

D5.2 Open benchmark dataset released
D5.3 Accuracy validation (300+ forms)
5 institutions in production

Integration Points

From WP4

Input	Description	Timeline
Field mappings	Source -> Target mappings	M16
CRM integration	API connection	M16
Confidence scores	Per-field certainty	M16

External Systems

System	Integration	Purpose
Wecan Comply	Primary	CRM platform
Partner CRMs	API	Field mapping
Document storage	S3/Azure	Input/output

Quality Assurance

Testing Levels

Level	Scope	Frequency
Unit tests	Individual components	Every commit
Integration tests	Component interaction	Daily
End-to-end tests	Full pipeline	Weekly
User acceptance	Real users	Monthly

Performance Targets

Metric	Target	Measurement
Form filling time	<30 seconds	Automated timing
UI responsiveness	<1 second	User perception
Batch processing	10+ docs/minute	Throughput test
Availability	99.5%	Uptime monitoring

Back to Work Packages

Previous: WP4