Vera · by Nex0

The AI agent whose decisions you can prove.

Vera resolves logistics and field-service disruptions the way your best operator would — except every decision is verified by a deterministic solver against your business rules, and every decision carries a complete audit trail. The LLM translates. It never decides.

In production with paying customers since 2025 · EU-hosted · Built in Finland

Decision tracetr_01J9XK

task “Truck 7 broke down on the E75 — reroute its afternoon deliveries.”

Why Vera exists

Two ways to automate operations. Both fail.

Pure-LLM agents are probabilistic. They cannot guarantee a constraint is never violated, and their “reasoning” is a narrative, not a proof. Nobody should dispatch a hazmat truck on a narrative.

Classical rule engines are rigid. Rules are configured by specialists, changes take weeks, and a free-text delay notice from a carrier never reaches them.

Vera combines both and discards the weaknesses. Plain-language rule capture from the LLM world; deterministic, provable decision-making from the symbolic world — in one agentic loop.

How it works

One pipeline. One guarantee.

Data and rules go in. A language model structures them. A solver decides. An agent acts. Everything becomes memory and audit trail.

01

Language in

Operators write rules and report disruptions in plain English. A language model translates them into structured, schema-validated facts and constraints — and that is all it does.

02

Decision computed

A deterministic Z3 constraint solver checks every candidate plan against every rule, names the exact rule any invalid option violates, and ranks the valid ones. Same input, same output — always.

03

Action with proof

The agent executes through REST and MCP at the autonomy level you configure — suggest, approve, or autonomous — and every decision is stored with its complete reasoning trace.

REST ingestionLLM + schema validationknowledge graph memoryZ3 solveragent (MCP)audit trail

What makes it different

Verifiable by construction, not explained after the fact.

Decisions are computed, not generated

Hallucination is structurally excluded from the decision path. The LLM translates language; the solver decides. A pure-LLM agent can't guarantee a hazmat load never crosses a residential zone. Vera can — mathematically.

Explanation is a proof, not a story

Every rejection names the exact rule and facts that caused it (the unsat core); every approval shows the constraints it satisfied. The trace is a permanent, compliance-grade audit record.

Rules in plain English

“Hazmat must not pass through residential zones.” Operators author and change rules themselves — schema-validated translation makes them machine-checkable. No consultants, no change-request queue.

Operational memory

Rules, facts and past decisions accumulate in a knowledge graph. Senior operators' knowledge becomes a durable, queryable company asset instead of leaving with retirement and turnover.

Configurable autonomy

Suggest, approve, or autonomous — per decision class, with human approval as the default. Autonomous actions are logged identically to approved ones.

Model-agnostic and sovereign

Runs on open-weight European models (Mistral, Teuken class, self-hosted via Ollama) or any API model you approve. Only prompt text reaches the endpoint; with self-hosting, nothing leaves your deployment.

Where it runs

Built for disruption-heavy operations.

Field service

Dispatch and emergency response

Technician out sick, refrigerant leak at a customer site, SLA clock running — Vera reassigns, reroutes and notifies within the rules for certifications, working time and response windows. In production with Finnish field-service operators since 2025.

3PL & logistics

Disruption handling at scale

A delayed truck, a closed terminal, a hazmat load that can't take the usual route. Vera evaluates options against all rules in seconds and either acts or proposes — operators stop reconstructing constraints from memory under time pressure.

Compliance

Audit trails that settle disputes

Every decision is recorded with the exact rules and facts behind it — directly usable in customer disputes, insurance claims and regulatory checks. Driving-time, ADR/hazmat and certification rules can't be silently bypassed.

For developers

Formal verification, one API call away.

The same engine that powers the console is a component your systems — and your agents — can call. REST for applications, MCP for agent frameworks. Send a plan, get SAT/UNSAT per rule and the unsat core when it fails.

POST /v1/plans/evaluate
{
  "feasible": false,
  "results": [{ "rule": "R2", "result": "violated" }],
  "unsat_core": {
    "rules": ["R2"],
    "facts": ["T-103.certifications",
              "WO-4811.equipment_class"],
    "explanation": "R2 requires a certified
      technician for gas-boiler work;
      T-103 holds no gas certification."
  }
}

Proof, not promises

Already running real operations.

Vera has been in daily production with Finnish field-service operators since 2025 — dispatch, routing and disruption handling, the same class of decision problem as logistics execution.

2025

in production with paying customers since

100%

of decisions carry a complete reasoning trace

Minutes

to resolve routine disruptions that used to wait on a senior dispatcher

0

decisions made by a language model

Customer identities are protected under our data processing agreements. Validation evidence — including the extraction evaluation suite and determinism test regime — is published in the documentation.

See your own rules verified in minutes.

Bring one real disruption from last week. We'll show you how Vera would have handled it — with the proof.