Build a Mobile Money Credit Scoring Agent with AI SDK and Qwen 3
In rapidly growing emerging markets across Africa, traditional credit bureau data is sparse or non-existent. Partnerships like Visa with _able and Onafriq demonstrate the push toward leveraging mobile wallet transaction histories, airtime purchases, and peer-to-peer transfer velocity to evaluate creditworthiness for underserved populations. In this challenge, you will construct an agentic credit scoring pipeline using the Vercel AI SDK integrated with the Qwen 3 model. Your system must ingest unstructured and semi-structured telecom/mobile wallet event streams, compute alternative risk indicators, and output a structured credit evaluation schema with verifiable reasoning chains.
What you are building
The core problem, expected build, and operating context for this challenge.
Develop an autonomous credit scoring pipeline using Vercel AI SDK and Qwen 3 to parse mobile money transaction logs into risk assessments.
How work is evaluated
Evaluates the accuracy of structured JSON output generation, feature extraction precision, and response latency.
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.
schema_conformance
Ensures output conforms strictly to the expected credit evaluation schema
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
score_precision
Mean Absolute Error between generated credit score and ground truth score • target: 5 • 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
Implement structured outputs with Vercel AI SDK streamObject and generateObject
Process non-traditional mobile money log formats into standardized financial telemetry
Integrate Qwen 3 model inference via OpenAI-compatible endpoints
Design zero-shot and few-shot prompts to minimize bias and hallucinated financial scores
Reference links and supporting material
Synthetic mobile money event dataset modeled on African wallet networks like Onafriq, including airtime top-ups, merchant payments, and agent cash-ins.
Challenge setup
Evaluates the accuracy of structured JSON output generation, feature extraction precision, and response latency.
Challenge input
JSON containing raw mobile money ledger events
Vercel AI SDK
Provides clean TypeScript primitives for structured LLM outputs and streaming.
Qwen 3
Cost-effective open model with strong multilingual and reasoning abilities.
Evaluated output
JSON containing credit_score, risk_category, and explanation
- Ensures output conforms strictly to the expected credit evaluation schema
- Mean Absolute Error between generated credit score and ground truth score • target: 5 • range: 0-100
- Benchmark: FinEval-AlternativeCredit
- Score Precision target: 5
- 1 public reference case
- JavaScript execution harness
- javascript 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
- Alibaba Cloud (Qwen) · Optional
- Qwen 2 · Optional
- Qwen 3 · Required
Evaluation contract
- schema_conformance · Weight 1
- score_precision · Weight 1
Metrics
- score_precision · Target 5
Benchmarks
- FinEval-AlternativeCredit
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing raw mobile money ledger events
Output: JSON containing credit_score, risk_category, and explanation
Execution harness: JavaScript
Evaluation tools
- Vercel AI SDK · Agent Framework
- Qwen 3 · Language Model
Recipe freeze marker
No recipe freeze time is recorded.