Jev 1.13 for business classification and routing
Where should you test Jev and small decision models? Explore ticket routing, document completeness, confidence thresholds and a Syntalith-led pilot.
Syntalith
A complaints queue contains delayed deliveries, incorrect invoices and damaged goods. Sometimes one message covers all three. Keyword rules become unreliable, while asking a large model for a lengthy explanation every time may add unnecessary work.
This is a sensible task for evaluating a decision model. Jev 1.13 on OpenRouter returns typed decisions with probabilities. The scenarios below are pilot proposals, not Jev deployments delivered to Syntalith clients.
Separate questions that lead to different actions
For a complaint, ask about the owning department, missing information and escalation needs. Each answer has a distinct role. The department determines the queue. Missing information can prepare a request for employee review. Escalation changes how the case is handled.
Allow for ambiguity. If several issues must map to one category, define priority or split the case. Without that rule, human reviewers may disagree about the correct answer too.
Where to start a pilot
| Process | Model task | Application or human checks |
|---|---|---|
| Help desk | Select queue and urgency signal | Available teams, permissions and exceptions |
| Order intake | Assess message completeness | Actual fields and ERP records |
| Service documentation | Suggest a fault category | Whether it fits the equipment and report |
| Internal knowledge search | Select a source to search | User access and document availability |
Start with an output that can be checked and reversed. Moving a ticket between queues is easier to assess than autonomously settling a claim. The automation scope should reflect the consequences of an error.
Set confidence thresholds from measurements
Suppose the system accepts decisions above a threshold and sends the rest to an operator. Raising that threshold may improve accepted-answer accuracy while reducing the share handled automatically. Measure both effects.
Consider an illustrative sample of 1,000 cases. The application accepts 600 decisions, including 12 errors. Coverage is 600 / 1,000 = 60%; accepted-decision accuracy is (600 - 12) / 600 = 98%. These numbers demonstrate reporting only. They are not Jev or Syntalith results. The remaining 400 cases still need work.
One average over the whole dataset would hide that distinction. The report should also identify difficult categories and the most expensive errors.
Test changes to business rules
A new complaints procedure can change the correct answer without changing the model. Version category definitions and expected behaviour. Include cases governed by both the previous and current rules, explicitly identifying the applicable policy.
A decision model may return a valid category, but the application still controls the resulting action. Before updating a CRM, it checks identity, authorization and data completeness. Logs should connect an outcome to the relevant model and rule version.
Jev or a local implementation?
Jev is a useful candidate where service access and processing terms fit the project. If data must stay inside a restricted network, we evaluate Kev, TinyJev or a small custom model. A shared response format does not establish equal quality.
Syntalith can prepare category tests, compare candidates and integrate the chosen version with exception handling. Describe the queue you want to improve, its volume and the current routing process. We can establish what to measure before a purchase decision.
Match a model to the task you need it to perform
Describe where your current AI falls short. We will compare model customization options, data requirements and the cost of running the resulting system.
Private LLMs and fine-tuning