Multilingual KYC Proof-of-Address Verification Engine with Claude Agents SDK
Cross-border payment expansion across Africa and emerging markets requires verifying identity and proof-of-address documents issued in diverse formats, languages, and character scripts (e.g. Arabic, French, Swahili, Amharic). Standard static OCR templates fail on non-standard utility bills. In this challenge, you will implement a multilingual proof-of-address verification agent using the Claude Agents SDK. The agent will analyze unstructured visual document scans, perform translation and entity extraction, and verify address details against customer registration databases.
What you are building
The core problem, expected build, and operating context for this challenge.
Develop a multilingual proof-of-address verification agent using Claude Agents SDK to parse non-standard global utility bills.
How work is evaluated
Evaluates address extraction accuracy and matching decisions across multilingual utility bill images.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
address_match_pass
Verifies match confidence is above 0.90 for valid documents
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
address_extraction_accuracy
Accuracy of extracted address string across global test samples • target: 95 • range: 0-100
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Initialize Claude Agents SDK with vision capabilities enabled
Parse unstructured multilingual utility bills (water, power, telecom)
Extract standardized street address, city, and date fields across French and Swahili inputs
Calculate fuzzy matching metrics against user-declared registration profiles
Reference links and supporting material
Synthetic global utility bill image samples in French, Swahili, and Arabic with corresponding ground-truth text annotations.
Challenge setup
Evaluates address extraction accuracy and matching decisions across multilingual utility bill images.
Challenge input
JSON containing document_image_base64 and user_declared_address
Claude Agents SDK
Anthropic's agent SDK leveraging Claude's multimodal vision features.
Evaluated output
JSON containing extracted_address, match_confidence, and onboarding_verdict
- Verifies match confidence is above 0.90 for valid documents
- Accuracy of extracted address string across global test samples • target: 95 • range: 0-100
- Benchmark: FinMultilingual-KYC-Bench
- Address Extraction Accuracy target: 95
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Claude Code · Required
- Claude Code · Required
- Claude 4.1 Opus · Optional
- Claude 4 Sonnet · Optional
Evaluation contract
- address_match_pass · Weight 1
- address_extraction_accuracy · Weight 1
Metrics
- address_extraction_accuracy · Target 95
Benchmarks
- FinMultilingual-KYC-Bench
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing document_image_base64 and user_declared_address
Output: JSON containing extracted_address, match_confidence, and onboarding_verdict
Execution harness: Python
Evaluation tools
- Claude Agents SDK · Agent Framework
Recipe freeze marker
No recipe freeze time is recorded.