There are definitely opportunities for improvement within AFS Entity Analytics. One suggestion is to lower the infrastructure barrier to entry; while the EC2, EMR, S3, Glue, or RDS setup is powerful, it can be heavy for teams without dedicated cloud data engineering support. A more managed or serverless deployment option would save significant setup time, especially for contracts where time is critical. Additionally, a faster cluster spin-up would help, as the current batch-oriented runs via EMR do not allow for quick ad hoc checks. More transparent cost visibility is another need; the EMR compute cost is separate from the license, and it is not always clear what the total operating costs are as data volume increases. Better estimation tools tied to expected data sizes would assist with budgeting for contract-based engagements. Usability and integration with other tools could be enhanced. For example, the UI is not particularly lightweight or analyst-friendly; most interaction happens through a custom dashboard that the pipeline outputs. The learning curve can be steep for non-engineers. The output format sometimes requires custom work to integrate downstream, particularly since the resolved entities and relationship data do not drop cleanly into the case management system, necessitating a transformation layer. The limited real-time integration options mean that hooking into systems expecting near-real signals, such as live fraud alerting, is infeasible without creating separate workarounds. The API surface appears secondary to the batch pipeline, with built-in integration patterns for event-driven use seeming afterthoughts. A dry run feature would be beneficial to test configuration changes or thresholds against a small sample before initiating a full EMR run. Another wish list item is the ability to save and switch between name configurations, such as strict KYC profiles versus a loose data deduplication profile, so settings would not have to be re-entered each time, enhancing user experience. Additionally, the documentation could provide more real-world examples, particularly concerning handling inconsistent legacy data, which would have made the early tuning phase smoother.
