TypeSafe AI Jev: 50 Financial Jobs, Verifiable Demos & Honest Limits

Replays. The entry's own caveats say the runs use synthetic or cached data rather than live requests, so the two hours show recorded outputs. Nothing there measures live latency, rate limits, error responses or degraded-endpoint behaviour, which in production finance work is most of the actual risk.

Not by itself. The copy says outputs, baselines and failures are saved, but no aggregate accuracy, precision/recall or false-positive rate is documented, and a failure log without labelled ground truth is not a rate. For AML or sanctions matching, missed detections matter far more than accuracy, and nothing published counts them.

The closing section is titled Fifty jobs, eight decisions, and that is as far as the page goes. It does not list the primitives, so you cannot tell how much of each workflow was Jev and how much was ordinary code. In a compliance context that ratio is the whole question, and it is left unstated.

Demo data, state and questions are hosted separately with code in a companion repository, which is promising, but the page documents no schema for that split, no pinned model version, and no statement that the saved outputs can be regenerated. Financial decisions deserve a rerun on your own data before you trust the pattern.

Accounting, payments, fraud, AML, crypto, portfolios, and market research. The entry describes fifty numbered demos closing with a section titled Fifty jobs, eight decisions.

Hosted separately, with code in a companion repository. The entry records that split, so the run and its supporting assets are not all in one place.

It is educational research, not financial advice, and the replays use synthetic or cached data rather than live requests. Nothing there measures live latency, rate limits or degraded-endpoint behaviour.