Juspay’s NeuroLink is a TypeScript toolkit for connecting applications to AI providers. Its classifier router can ask a decision model, including Jev, which kind of model a request needs.
Make the risks asymmetric
- One request asks about task difficulty, tool and vision needs, risk, and conversation scope. When multiple candidates are available, it also asks which model to use.
- The code defaults to a confidence bar of 0.3 for an upgrade and 0.6 for a downgrade from the neutral difficulty tier. These are project settings, not measured accuracy rates.
- A risk probability above 0.7 can force the difficulty to at least hard. That override can raise the tier, not lower it.
- Missing or weak difficulty answers fall back to a heuristic. The router also catches decision-call errors and uses the heuristic.
Protect the boundary
The classifier sends the request text and selected metadata to the decision provider. The code leaves the opaque session ID out of the decision state, but the prompt itself can still contain sensitive information.
Confidence describes the model’s output, not whether its routing choice is correct. Evaluate thresholds and end-to-end cost on labeled requests, and do not treat the risk classifier as a security boundary.
Caveat: The source was inspected, not run. The project’s routing quality, calibration and performance were not independently reproduced.