Your agent closed the ticket — and invented a discount to do it.
Axon listens to every voice and chat conversation, gives an A–F grade per dimension and shows where the agent failed, why, and how to fix it. From customer-facing voice and chat agents to RAG copilots and tool-using agents — AI quality stopped being guesswork.
“I can apply a 20% discount right now.” — unauthorized concession, invented to avoid churn.
The dashboard is green. The customer isn't.
Observability tells you what happened — latency, cost, volume. None of it reveals whether the agent lied, went off-script or promised something that doesn't exist. Axon grades the content of the conversation: real quality.
What happened
Latency, tokens, cost, handoff rate. Machine metrics — silent about the experience.
Why it happened — and what to do
Resolution, anti-hallucination, adherence, tone. The exact failure, the evidence and the fix.
It doesn't stop at voice and chat.
If it's an AI agent in production and a bad decision is expensive, Axon grades it. The rubric adapts to the type of agent — the A–F grade stays the same language for the whole leadership.
Voice & chat agents
Resolution, anti-hallucination, script adherence, tone and safety.
Configurable rubric
From a voice agent to a RAG copilot, the rubric adapts — the A–F grade stays the same language for leadership.
Have an AI system that isn't on the list? The rubric is configurable — talk to us.
From connection to fix in four steps.
Connect the agent
Upload transcripts or stream via API. Voice, chat or WhatsApp.
Get the report
An A–F grade per dimension, with confidence level and sample size.
See the evidence
The exact turn that failed, highlighted, with the explanation alongside.
Lock it in regression
Turn the fix into a test. The fixed problem doesn't come back — gate on deploy.
A verdict leadership reads in seconds.
The overall grade sums up the health. The six dimensions show where it hurts. The color — green, amber, red — speaks before you even read.
Everything between “looks fine” and “under control”.
Evidence at the exact turn
Stop reading the whole transcript. Axon points to the turn that failed and explains why.
Recommended fix
Root cause plus prompt patch, guardrail and eval — ready to apply, not just a diagnosis.
Agent trajectory & RAG
Beyond the final answer: did it call the right tools, in the right order, and ground the response in the retrieved context (faithfulness)?
Regression for agents
Golden sets that run on every deploy and block the version that regresses. Real CI for AI.
Trust the judge
Calibrate the AI grader against your own human reviewers and track agreement (Cohen's kappa). No black box.
Degradation alerts
Know before the customer complains. A grade drop becomes an alert on Slack, email or webhook.
Built for people who ship to production.
Start with one call — or pnpm add @gabrielonrails/axon-sdk. Typed SDK, REST, webhooks and versioned rubrics. No magic, no black box.
$ pnpm add @gabrielonrails/axon-sdk
import { Axon } from "@gabrielonrails/axon-sdk";
const axon = new Axon({ apiKey: process.env.AXON_API_KEY! });
// Envia uma transcrição — idempotente por `id`, avaliada em segundo plano
await axon.ingest("cobranca", [{
id: "conv-001",
channel: "voz",
turns: [
{ role: "client", text: "Quero um desconto." },
{ role: "agent", text: "Posso aplicar 20% agora." },
],
}]);
// Aguarda a avaliação e lê o boletim A–F
const status = await axon.waitForEvaluation("cobranca", "conv-001");
const report = await axon.getReport("cobranca");
console.log(report.overall, report.dimensions);
Quickstart
From zero to your first report in ~10 minutes.
ReadNode SDK
pnpm add @gabrielonrails/axon-sdk — typed, idempotent, with built-in retries.
ReadIngestion API
Send voice and chat transcripts, in batch or streaming.
ReadRead results
getReport, getConversation and waitForEvaluation pull the A–F report.
ReadIdempotency & retries
Re-send by id without duplicates; automatic exponential backoff.
ReadGrade reference
How A–F is calculated, dimension by dimension.
ReadUnlimited evaluations, on your own AI key.
There's no catch in “no evaluation cap”. The judge runs on the AI provider key you plug in — your contract, your cost, your data path. Axon doesn't resell inference or mark it up.
Cost control
You see and own the inference spend. FinOps shows cost per agent, model and day, with a monthly budget alert.
Data path you control
PII is redacted before the judge ever sees it, and the transcript goes to your provider under your terms.
Any provider
Anthropic, OpenAI or any OpenAI-compatible endpoint (Gemini, Groq, OpenRouter, local).
Pricing
Plans sized to how many agents you run
Pay monthly via PIX. Pick a plan to create your account and activate.
Boletim
R$ 97,00/mo
For one agent in production, run by you.
1 agent, no evaluation cap, ready-made recommended fix, degradation alerts and full history. The whole product, sized for a solo operator.
Pro
R$ 397,00/mo
For when you've shipped and want to sleep at night.
Up to 5 agents, everything in the micro-SaaS, plus basic regression and more alert destinations. The small-team plan with more than one agent live.
Team
R$ 1.997,00/mo
For many agents and constant deploys.
Everything in Pro, plus voice, chat and WhatsApp in the same report, full regression/CI with golden sets that block regressing deploys, multiple projects and configurable rubrics.
Enterprise
Custom
For AI at scale with governance demands.
Everything in Team, plus SSO and RBAC, data residency and LGPD, white-label, SLA, dedicated rubric and specialist support. Annual contract, starting with a paid diagnostic (R$ 25k–60k) on real transcripts, creditable to the contract.
Want to try before subscribing? Pay R$5, send one conversation, get one full A–F report.
Test 1 report for R$5Monthly recurring PIX (Woovi). Cancel anytime.
Questions, answered.
How is Axon different from observability tools?
Observability tells you what happened — latency, cost, volume. Axon evaluates the content: it grades each conversation A–F across dimensions, points to the exact turn that failed, explains why, and gives a concrete fix. It's the quality layer, not another log.
Why do I bring my own AI key (BYOK)?
The judge runs on your own Anthropic, OpenAI or compatible key — so your data never trains a third-party model, you control cost, and you can even judge across model families to reduce bias. No managed key, no lock-in.
Is my conversation data safe?
PII (emails, documents, cards, phones, names) is redacted before anything is stored or sent to the judge. You set a data-retention window, can delete all your data anytime (LGPD), and the judge only ever sees redacted text.
Which agents and channels does it cover?
Voice, chat and WhatsApp — from customer-facing agents to RAG copilots and tool-using agents. It also evaluates tool calls (agent trajectory) and retrieved context (RAG faithfulness).
How do I integrate?
Upload transcripts (JSON/JSONL) or send production traces via the REST API or the typed SDK (pnpm add @gabrielonrails/axon-sdk). Ingestion is idempotent and evaluation runs automatically.
Can I trust the AI judge's grades?
Every problem cites the literal turn (we verify the quote actually exists), confidence reflects agreement across samples, and you can calibrate the judge against your own human reviewers with Cohen's kappa. No black box.
How much does it cost?
Plans from R$97/mo up to Team, billed via Pix Automático. Want to try first? A R$5 pack runs one full scorecard on your real data — pay-before-use, no commitment.
Connect your first agent today.
See the real quality of your agents before the customer does.
Already have an account? Sign in