Igniting the core
Human-in-the-Loop · Machine learning at the edge · Multi-tenant
AI proposes, scores and executes only what it is allowed to. People approve everything critical. A real automation engine, predictive scoring trained on your own closed deals, and an immutable audit trail for every move.
18
integrated operational modules2.5pp
model calibration error8
explainable factors per prediction100%
critical actions audited300+
edge deployment citiesThree capabilities that set NEXUS CORE apart from any CRM with “AI” bolted on top.
Real rules: trigger → conditions → actions that write to the CRM. Dry-run simulation showing why each condition matched, an action budget, per-lead cooldown to prevent loops, and a mode governed by human approval.
Dry-run plus full traceabilityAn 8-factor logistic regression trained on your closed deals inside the Cloudflare runtime. L2 regularization anchored to a calibrated domain prior, so it is accurate from day one and keeps learning as you close.
Calibration error of 2.5 pointsEvery prediction breaks down into the exact contribution of each variable: −10.2pp for inactivity, +1.4pp for a high score. No black boxes: your sales team understands the number and can argue with it.
Zero opaque decisionsEvery operation is classified as low, medium, high or critical. Financial, legal and irreversible actions never run on their own: they enter an approval queue with full context so a person decides.
Human approval mandatoryWho, what, when, with which data and under which rule. All recorded, versioned and reversible. If something ran, the exact trace of why it ran exists.
Complete permanent recordOrganization isolation is not an application-level filter: it lives in the schema. Five roles with a granular permission matrix across 18 modules, verified on every server endpoint.
401 and 403 verified end-to-endThe real risk of AI inside a company is not that it makes mistakes: it is that it acts without anyone knowing. Here that is structurally impossible.
It analyzes the pipeline, scores leads, drafts copy, flags deals at risk and stages actions 24/7.
Every action receives a label: low, medium, high or critical. The label defines who is allowed to execute it.
Critical items enter an approval queue with full context. Approve, reject or block.
An immutable record of the decision, the actor, the timestamp and the exact data. Reversible.
Turn on what you need today and switch on the rest as you grow. Every module honors the same permission matrix.
Leads, pipeline, funnel, accounts, tasks, contracts and analytics on a multi-tenant model.
System coreVisual builder for executable rules with simulation, budget and cooldown.
Executes for realConversion scoring with our own model trained at the edge and full explainability.
Learns from your winsOutreach sequences across email, call and task with templates per industry.
Nothing goes coldWeighted pipeline, deal velocity, cohorts and revenue projection.
Defensible forecastQueue of critical actions with context, history and a traceable decision.
The human gateTriggers, conditions, retries and escalation with approval gates.
No-code processesAI skills versioned per vertical: legal, accounting, medical, real estate.
AI that truly understandsAutonomous agents with limited scope, budget and activity reporting.
Bounded autonomyForm and screen builder with preview, validation and publishing.
No IT dependencyOptimization for classic and AI search engines: entities, FAQs and schema.
Real visibilityRooms with recording, transcription, AI summary and downstream automation.
With follow-upInbound and outbound with sentiment, dynamic scripts and human escalation.
Always answeringAPI keys, HMAC-signed webhooks, email, storage and external CRMs.
Everything connectedConfigurable reports with export and metrics derived from the real pipeline.
Data, not opinionsImmutable log queryable by actor, action, module and time window.
Compliance readyNo servers to manage. Tens-of-milliseconds latency from any continent.
Code runs on the node closest to each user. No cold starts, no infrastructure to maintain.
A framework under 14 kB with strict end-to-end typing. The entire bundle weighs less than a single cover image.
Real relational persistence with multi-tenant isolation in the schema and migrations versioned in the repository.
Five roles verified server-side across 18 modules. No endpoint trusts the frontend.
Scoring is validated against the Bayes ceiling, not an arbitrary threshold. Calibration, learning, regularization and monotonicity, all in a runnable test bench.
// L2 regularization anchored to the domain prior.
// It does not push weights toward zero: it anchors them to
// expert knowledge, so the model is never worse than the prior.
for (let j = 0; j < DIM; j++) {
const pull = l2 * (w[j] - PRIOR_WEIGHTS[j]);
w[j] -= lr * (grad[j] / n + pull);
}
// Blend proportional to sample size:
// with little data the prior dominates, with plenty the fit does.
const alpha = Math.min(1, n / MIN_SAMPLES_TRUST);
const final = w.map((wi, j) =>
alpha * wi + (1 - alpha) * PRIOR_WEIGHTS[j]
);
// And every prediction decomposes into factors:
return {
probability: sigmoid(logit),
contributions: features.map((v, j) => ({
label: FEATURE_NAMES[j],
impact: w[j] * v // ← explainable
})).sort((a, b) => Math.abs(b.impact) - Math.abs(a.impact))
};The model is not hardwired to a vertical: it learns from your operational history. Fields, stages and cadences adapt; the decision engine stays the same.
No forced contracts. No surprises on the AI bill: cost control lives inside the product.
For teams getting started
$49/user/month
For teams already selling at scale
$149/user/month
For operations that cannot fail
Custom
Every action the AI proposes is classified by risk. Low-risk actions execute on their own; financial, legal or irreversible ones sit in an approval queue and only execute once a person with the right role authorizes them. Everything lands in an immutable audit log.
An 8-factor logistic regression trains directly at the edge on your closed deals. Until there is enough data it operates with a domain prior calibrated against your per-stage base rates, with a calibration error below 2.5 percentage points. Every prediction includes the exact contribution of each factor.
Yes. The engine has a simulation mode: it shows which leads match, why each condition matched and which actions would execute, without writing anything to the database. There is also an action budget and a per-lead cooldown to prevent loops.
It runs on Cloudflare Workers across more than 300 cities, with Hono as the framework and PostgreSQL over a REST API for persistence. There are no servers to manage and latency is measured in tens of milliseconds from anywhere in the world.
The platform is universal by design and AI skills are versionable per vertical: accounting, legal, medical, real estate, education, logistics and more. The data model is multi-tenant, so every organization keeps its own isolated data.
The model starts from an already calibrated domain prior, so it predicts correctly from the very first lead. As you close deals the weight shifts toward your real data in proportion to sample size: there is never an abrupt jump or a blind period.
Step into the full platform with real demo data. Nothing to install, no credit card.