Kafene: Financial Operations Platform & Automated Data Quality Suite
A unified admin platform and Spark-based data quality suite for lease-to-own finance, replacing multi-portal application review with one dashboard that flags bad third-party data before underwriting.
The Problem
Operations teams reviewing high volumes of consumer finance applications had to juggle three systems per application: the Plaid portal for banking data, the Scienaptic portal for credit scoring, and an internal CRM for everything else. Every review meant re-finding the same applicant across portals, and throughput scaled only with headcount.
Worse, nothing watched the data itself. Third-party feeds arrive as deeply nested JSON, and when banking or credit fields came back missing or suspicious, nobody knew until after an underwriting decision had already been made on incomplete information. The only safeguard was analysts running manual SQL spot-checks—sporadic, slow, and easy to skip under volume.
What We Built
We built a unified admin dashboard that aggregates Plaid banking data, Scienaptic credit scores, and the internal CRM into a single React interface, backed by a Play Framework API that pulls from PostgreSQL, S3, and third-party sources—with fuzzy search, lazy loading for heavy third-party payloads, Redis caching for frequently accessed applications, and role-based access for operations teams and managers.
Underneath it runs an automated data quality layer: PySpark jobs flatten the nested JSON from financial APIs, explode arrays, and compute null, filled, and unique metrics across 10,000+ files daily, generating per-application and aggregated summary metrics written to S3. Color-coded indicators surface incomplete or suspicious data directly in the review interface, so quality problems appear before underwriting instead of after.
Single-pane application review
Banking data, credit scores, and CRM records for an application appear in one dashboard, eliminating the portal-switching that dominated every review.
Automated daily quality profiling
Spark jobs profile every nested JSON dataset daily—null rates, fill rates, uniqueness—replacing the manual SQL spot-checks analysts used to run by hand.
Visual data quality indicators
Color-coded flags in the review UI mark missing or suspicious third-party data per application, so reviewers see the reliability of the data alongside the data itself.
Performance-tuned review workflow
Fuzzy smart search, lazy-loaded third-party payloads, and Redis caching keep the dashboard responsive even with heavy multi-source data behind every application.
How It Works in Practice
An operations reviewer working through the morning queue of lease-to-own applications.
- 1
Find the application
Smart search brings up the applicant instantly; the dashboard assembles CRM, banking, and credit data in one view instead of three portals.
- 2
Check the quality flags
Color-coded indicators—computed by the overnight Spark profiling run—immediately show that the banking payload is complete but a credit field looks suspicious.
- 3
Drill into the data
Lazy-loaded third-party detail confirms whether the flagged field is a data problem or a real signal, without the reviewer leaving the interface.
- 4
Decide with confidence
The application proceeds to underwriting with known-good data, or gets routed back for re-pull—before a decision is made on bad inputs.
Engineering Approach
Review time is the bottleneck
The design target was minutes saved per application at volume—one interface, instant search, pre-fetched data—because operations throughput compounds from small per-review costs.
Quality as a pipeline, not an audit
Data quality runs as scheduled Spark jobs producing metrics per application and in aggregate, making "can we trust this feed today" a dashboard fact rather than a quarterly investigation.
One platform, three layers
The React UI, Play Framework API, and PySpark profiling suite were built as one coherent system, so quality metrics flow straight into the review workflow instead of living in a separate tool.
Have a similar problem?
Tell us what you're working on and we'll tell you honestly whether we can help.
