GO LOCAL.AI
Menu

Routing that changes
with you.

The small routing model is trained on your real work and then keeps learning and improving in production so you can keep saving as your AI use changes.

Get started Routing Engine
Customer-specific classifier training Customer-specific workload and failure examples are anonymised and labelled. A candidate classifier is fine-tuned, evaluated and shadow-tested against quality and privacy requirements. A passing candidate is released automatically when enabled by the customer, or after approval; failed candidates return to training. Production feedback feeds the next cycle. CUSTOMER-CONTROLLED DATA BOUNDARY Workload + failure examples Anonymise · label · map data boundaries Fine-tune candidate Customer-specific examples Evaluate + shadow test Quality + routing + privacy Below threshold: refine and retrain PASS + RELEASE Release the update Approved classifier + rollback Production feedback Failures · drift · unknown work Customer-specific classifier training Customer-specific workload and failure examples are anonymised and labelled. A candidate classifier is fine-tuned, evaluated and shadow-tested against quality and privacy requirements. A passing candidate is released automatically when enabled by the customer, or after approval; failed candidates return to training. Production feedback feeds the next cycle. CUSTOMER-CONTROLLED TRAINING Workload + failure examples Anonymise · clean · label Map privacy constraints Fine-tune candidate Customer-specific examples Evaluate + shadow test Quality + routing + privacy Below threshold: refine and retrain PASS + RELEASE Release the update Approved classifier + rollback Production feedback Failures · drift · unknown work LEARN FROM THE WORK. TEST EVERY UPDATE.

Trained on what good outcomes mean for you.

Training input What it teaches the classifier
Your workloads Anonymised, structured examples of the tasks, prompts and workflows it will see in production.
Benchmark outcomes Which model destinations meet the quality requirement for each type of work, and which suitable options cost less.
Data Boundaries Defined sensitive or risky data categories, approved destinations and work that must remain private.
Production feedback Quality issues, misroutes, wrong escalations, low-confidence decisions and unknown workloads.

The self-improvement
cycle.

  1. Monitor performance and  misroutes.

    The Monitoring Engine keeps gateway logs of requests/prompts associated with poor quality, failed routes, wrong escalations, low confidence or unfamiliar work. 

  2. Prepare the next training set.

    The Assessment Engine automatically anonymises, cleans and deduplicates collected logs and data. It labels each example by task, data category and suitable outcome.

  3. Train a new classifier candidate automatically.

    The Routing Engine creates an improved version of the classifier model on the updated dataset. This changes the routing model, not the weights of the cloud or open-weight models serving requests.

  4. Evaluate the decisions.

    The Platform compares the candidate with the current classifier against the agreed workload and quality criteria. It tests known work, new examples, sensitive categories and escalations. Refine and retest candidates that miss the threshold.

  5. Shadow-test before release.

    The Routing Engine runs the candidate alongside the existing routing setup without letting it control production destinations. Compare its decisions on representative traffic, including unknown and sensitive workloads.

  6. Release and keep learning.

    After evaluation and shadow testing, the Platform can deploy automatically if enabled or approve each update. It keeps the previous version ready for rollback.

Maintain your standards.

 View release checks
Release check What has to hold
Routing and task quality The candidate meets the configured success threshold on the work it routes, without treating lower cost as proof of quality.
Sensitive-data handling Client-defined sensitive categories stay within approved private routes, including in evaluation and shadow testing.
Unknown and difficult work Low-confidence decisions and escalations follow the agreed safe behaviour. When no compliant destination exists, the request is blocked.
Production approval After checks pass, deploy automatically if enabled or require approval. Rollback remains available.
Routing decisions in production Infrastructure recommendations Security and training data

Find out what your AI
should be costing you.

Book a 30-minute review Email us