Route the tool call.
Score tools against an incoming request before the next stage of your agent runs. Keep execution and permissions in your application.
Evaluate tool routing ↗PURPOSE-BUILT MODELS. FINITE CHOICES.Your next tool call
Your next tool call
starts with
a decision.
Give ZEV the context and the choices. Get a score for every candidate in one forward pass, inside your infrastructure.
Rust / Candle · PyTorch · Inference on your infrastructure
CHOICE ISOLATION01 / ROUTING
INCOMING REQUEST
“I can’t sign in to my account.”
CANDIDATES / ILLUSTRATIVE SCORES
01 Billing & invoices 0.5
02 Account access 2.4
03 Technical support −0.7
Same choices. Same scores.
Architecture illustration · example values, not model predictions. Reordering moves each score with its candidate. See the guarantee →
One forward passScores over the choices you supply.
Order-equivariant by designCandidate order does not define its score.*
Runs on your infrastructureIntegrate ZEV directly into your application.
*For unchanged input text and weights, subject to numerical tolerances. Details & tie handling
01 / THE RIGHT KIND OF PROBLEMYou already know
You already know
the possible actions.
Your application has a tool catalog, a set of queues, or a defined workflow. ZEV evaluates those candidates against the supplied context and question.
Find the right queue.
Evaluate support queues or incident destinations against a ticket. Measure costly misroutes and define an escalation path.
Choose the next branch.
Score a bounded set of workflow actions. Validate behavior on your domain before connecting predictions to production.
Evaluate workflows ↗02 / INSIDE THE DECISIONShared context.
Shared context.
Independent choices.
Each candidate reads the context and itself. The attention graph blocks communication between competing choices.
01 / YOUR APPLICATIONContext + choicesA question with possible actions
02 / ZEVScore each candidateOne non-autoregressive pass
03 / YOUR APPLICATIONApply your policyExecute, escalate, or review
03 / PROOF BEFORE PRODUCTIONA clear guarantee.
A clear guarantee.
A measurable fit.
Order equivariance is an architectural property. Predictive accuracy, tool selection, and sub-30ms latency are validated across public enterprise benchmarks.
01
Define success.
Agree on representative examples, acceptable errors, deployment constraints, and the current baseline.
02
Measure the decision.
Evaluate predictive quality, fallback behavior, order sensitivity, and latency on the target hardware.
03
Make the commercial call.
Use the results to choose an integration scope and a commercial license.
04 / THE BUSINESS CASEPut your numbers
Put your numbers
through it.
Explore the economics of moving a portion of your routing decisions to a dedicated decision model. Change every assumption to reflect your business.
USD · Editable assumptions, not measured outcomes. The license default is a planning figure from the commercial draft, not a quote. Include your support and operating overhead in your operating costs.
ESTIMATED FIRST-YEAR DIFFERENCE
Potential savings against your current routing cost.
- Current annual cost
- $24,000
- Remaining routing cost
- $7,200
- Compute + license
- $8,700
- One-time integration
- $3,000
Offloaded decisions avoid the current decision cost. All other decisions retain it. Downstream generation costs are excluded.
Validate the assumptions ↗05 / BEFORE YOU ASKBuilt for a specific
Built for a specific
kind of decision.
A few things to know before you evaluate.
Does ZEV replace my whole agent?
ZEV scores supplied choices. Your agent still handles planning, generation, tool execution, and permissions. Evaluate ZEV for a bounded decision step within that system.
Can it run without an external model API?
The Rust and PyTorch implementations support inference on your infrastructure using model weights and a tokenizer. Your surrounding deployment determines network access and operational controls.
What if none of the choices is right?
You can include and train an explicit “none” or escalation candidate. Automatic abstention is not an architectural guarantee. Confidence thresholds need domain-specific evaluation.
Can we train on our own examples?
The repository includes training workflows and a local workspace for CSV intake, training jobs, and local HTTP deployment. Commercial training rights and adaptation support depend on the agreed license.
Can we buy a hosted API today?
The documented workspace currently runs locally. Hosted infrastructure and billing integrations remain deferred. Start with an evaluation for deployment on your infrastructure or embedding.
YOUR WORKLOAD. YOUR ACCEPTANCE CRITERIA.
Evaluate ZEV Make the next decision on evidence.
Bring a routing workload. Find out where ZEV fits.