ChevLink AI

Intelligence built
for security at scale

Purpose-built AI models powering every ChevLink product. Each model does one job, runs in milliseconds, and requires no GPU or cloud dependency.

6
AI systems
24
Specialised models
.onnx
Runtime format
24/7
Continuous training
Live

Running in production

Actively deployed across ChevLink's bots and dashboard.

Live v1.2 LLM
Nova
ChevLink's AI security assistant
Conversational assistant built into the dashboard. Manages Aegrix and Cyber settings, answers security questions, calls specialised AI models as tools, and navigates the platform on your behalf.
Phi-4 Mini base 3.8B params 2.5 GB Tool calling
Nova 1.3 — coming next
LoRA fine-tuned on real ChevLink interactionsPlanned
Improved tool-call accuracyPlanned
Live v1.1 ONNX
CyberAI
Powers the Cyber Discord bot
A suite of fast ONNX classifiers and anomaly detectors running inside the Cyber bot. Detects raids at onset, scores new accounts for risk, filters spam, and matches ban reasons across the CRN network.
4 models Runs in-bot <5ms inference CPU only
Included models
Join Spike DetectorLSTM autoencoder
Account Risk ScorerXGBoost
Message Spam ClassifierTF-IDF + LogReg
Ban Sync MatcherCosine similarity
Live v1.1 ONNX
AegrixAI
Powers the Aegrix Discord bot
Aegrix's intelligence layer. Classifies threat signals, scores unverified users at the gate, predicts auto-lockdown triggers, and suppresses low-confidence alerts before they reach admins.
4 models Runs in-bot <5ms inference CPU only
Included models
Threat Level Classifier3-class GBT
Verification Risk ScorerLogReg
Lockdown Trigger PredictorLSTM
False Positive ReducerRandom Forest
Coming Soon

In development

Being trained and validated before production deployment.

Coming soon v1.1 ONNX
PulseAI
Powers ChevLink Pulse monitoring
Uptime anomaly detection, alert severity ranking, and notification fatigue filtering. Ensures the right alerts reach the right people without noise.
3 modelsLSTM autoencoderAlert ranking
Included models
Alert Severity RankerGBT regression
Uptime Anomaly DetectorLSTM autoencoder
Notification Fatigue FilterLogReg
Coming soon v1.1 ONNX
HorizonAI
Threat intelligence & risk forecasting
ChevLink's risk intelligence layer. Forecasts community threat levels 24h ahead, links threat actors across platforms, classifies intelligence signals, and detects unusual risk spikes.
4 modelsCross-platformNova tool calls
Included models
Threat Level ForecasterLSTM regression
Cross-Platform Threat LinkerEmbeddings
Intel Feed Classifier5-class GBT
Risk Trend Anomaly DetectorIsolation Forest
Coming soon v1.1 ONNX
Nova Tool Models
Nova's infrastructure intelligence
Small ONNX models Nova calls as tools mid-conversation. Detects suspicious activity on Vultr and AWS, scores login events for credential attacks, and classifies infrastructure anomaly types.
5 modelsVultr · AWSIn-context tool calls
Included models
Vultr Anomaly DetectorLSTM autoencoder
AWS Anomaly DetectorLSTM autoencoder
Infra Threat Classifier5-class GBT
Login Anomaly ScorerXGBoost
API Abuse DetectorLSTM autoencoder
AutoTrainer

Models that improve themselves

A continuous training pipeline — no human intervention required.

ChevLink AutoTrainer
Every inference is logged. Every prediction tracked against real outcomes. AutoTrainer retrains each model nightly, A/B tests the challenger against the champion, and promotes automatically when accuracy improves. Rollback triggers if performance regresses.
Data collection Scheduled retraining A/B evaluation Auto-promotion Nova LoRA fine-tuning Version bumping Auto rollback
Roadmap

What's next

Nova and the ONNX suite are versioned independently.

Nova 1.2 — Phi-4 Mini
Upgraded from Qwen 2.5. Better instruction following, native tool calling, runs on a $24/mo Vultr VPS without a GPU.
Released
CyberAI 1.1 & AegrixAI 1.1
First-generation ONNX model suites deployed into the Cyber and Aegrix Discord bots.
Released
PulseAI 1.1 & HorizonAI 1.1
Alert ranking, uptime anomaly detection, cross-platform threat linking, and 24-hour risk forecasting. Horizon becomes the shared intelligence layer.
In development
AutoTrainer launch
All ONNX models begin self-improving nightly. Nova starts LoRA fine-tuning on real dashboard conversations.
Planned
Nova 1.3 — First fine-tune
LoRA fine-tuned on real ChevLink interactions. Shorter answers, better tool accuracy, ChevLink-specific knowledge at the weight level — not just the prompt.
Planned
Million-parameter ONNX upgrade
All ONNX models upgraded from sklearn baselines to million-parameter PyTorch neural networks. Deeper LSTMs, wider hidden layers, attention on time-series models.
Planned
Nova 2.0
Major version. New base model or full architecture change. Full ownership of model weights — ChevLink's own LLM trained on security domain data.
Future