Getting started

MLOps.dev Documentation

From curl install to your first edge deployment in under 10 minutes. No Kubernetes. No YAML manifests. Just a 7.4MB binary and a CLI.

✓

Early access: These docs cover the current private beta. Join the waitlist to get access. Questions? Email hello@mlops.dev.

1. Install the agent

Run this on your edge device. Works on any Linux ARM or x86-64 system with 256MB+ RAM.

BASH

# One-line install
curl -fsSL get.mlops.dev | sh

# Verify installation
mlops --version
# mlops-agent v0.6.0 linux/arm64

# Check device is registered
mlops status

ℹ

The install script detects your architecture (arm64, armv7, amd64) and downloads the correct binary. The agent binary is 7.4MB, statically compiled with musl libc — no runtime dependencies.

2. Push your first model

From your development machine, register a model and deploy it to your device.

BASH

# Authenticate
mlops login

# Register a model (ONNX, TFLite, or TensorRT)
mlops push ./model.onnx --name defect-detector --tag v1.0

# Deploy to your device
mlops deploy defect-detector:v1.0 --target device-id-here

# Check deployment status
mlops status --fleet

PYTHON SDK

import mlops

client = mlops.Client(api_key="your-api-key")

# Register model
model = client.models.push(
    path="./model.onnx",
    name="defect-detector",
    tag="v1.0"
)

# Deploy
deployment = client.deploy(
    model="defect-detector:v1.0",
    target="jetson-prod-01"
)
print(deployment.status)

The edge agent

The mlops-agent is a single statically-compiled Go binary that runs on your edge device. It owns the deployment loop — the control plane signals desired state, the agent reconciles.

Property	Value
Binary size	7.4MB (arm64 production build)
RAM usage	~14MB at steady state
Startup time	<200ms
Architectures	arm64, armv7, amd64
OS	Any Linux, kernel 4.9+
License	Apache 2.0
Source	github.com/Raghunath2604/Raghunath2604-mlops-dev/Raghunath2604-mlops-dev

Offline-first sync

The agent uses a pull-based idempotent sync protocol. It polls the control plane for desired state every 30 seconds and reconciles itself. During outages, it buffers telemetry locally in SQLite and flushes on reconnect.

→ Tested: 6-hour outage on 2G connection. Zero data loss.
→ Delta compression: 95.7% smaller update packages (bsdiff).
→ Buffer cap: 500MB local SQLite, discards oldest records when full.

Drift detection

Statistical drift detection runs on the device itself using KL divergence on the input distribution, compared against a baseline captured at deployment time.

AGENT CONFIG (mlops.yaml)

drift:
  enabled: true
  method: kl_divergence
  window_size: 200        # inferences in rolling window
  check_every: 100        # recompute every N inferences
  warn_threshold: 0.4    # KL divergence warning level
  alert_threshold: 0.7   # KL divergence alert level
  alert_webhook: ""       # optional: Slack/PagerDuty URL
  monitors:
    - input_distribution
    - output_confidence

✓

KL drift detection adds ~60 microseconds average overhead per inference at 30fps. On a Jetson Nano, the full computation takes under 2ms.

Canary deployments

Stage model updates across your fleet by hardware class. Health gates automatically halt rollout if accuracy drops or latency increases beyond thresholds.

BASH

# Staged rollout by hardware class
mlops deploy defect-detector:v2.0   --stage "hw_class=jetson_orin,count=1"   --stage "hw_class=jetson_orin,pct=100"   --stage "hw_class=jetson_nano,pct=25"   --stage "hw_class=all,pct=100"   --health-gate "accuracy_delta=-0.03"   --health-gate "latency_delta=+0.20"   --stage-interval 30m

# Monitor rollout
mlops rollout status

# Manual rollback if needed
mlops rollback --to defect-detector:v1.0

CLI reference

Command	Description
mlops login	Authenticate with the control plane
mlops push <file>	Register a model file in the registry
mlops deploy <model:tag>	Deploy a model version to fleet
mlops rollback	Roll back fleet or device to previous version
mlops status	Show fleet health, device list, model versions
mlops devices list	List all registered devices
mlops drift	Show drift scores across fleet
mlops logs <device-id>	Stream device event log
mlops config set	Update agent configuration

Agent configuration

/etc/mlops/mlops.yaml

agent:
  device_id: ""              # auto-generated on first run
  api_endpoint: "https://api.mlops.dev"
  api_key: ""               # from dashboard → Settings
  heartbeat_interval: 30s
  sync_interval: 30s
  telemetry_buffer_mb: 500

models:
  runtime: auto              # auto | onnx | tflite | tensorrt
  model_dir: /opt/mlops/models

drift:
  enabled: true
  warn_threshold: 0.4
  alert_threshold: 0.7

logging:
  level: info               # debug | info | warn | error
  output: journald           # journald | file | stdout

Supported hardware

Device	Class tag	Notes
NVIDIA Jetson Orin NX/AGX	jetson_orin	TensorRT INT8 recommended
NVIDIA Jetson Nano	jetson_nano	TFLite FP32 or ONNX
Raspberry Pi 5 / 4B	rpi5 / rpi4	TFLite INT8 recommended
Raspberry Pi Zero 2W	rpi_zero	Minimal mode, 6MB agent
Google Coral Dev Board	coral	Edge TPU delegate
x86-64 industrial PC	x86_64	Full feature set
Custom ARM Cortex-A	arm_custom	Requires Linux kernel 4.9+

Model formats

Format	Best for	Latency (Jetson Orin)
TensorRT INT8	Jetson GPU — maximum throughput	3.8ms · 263fps
TensorRT FP16	Jetson GPU — accuracy-critical	6.1ms · 163fps
ONNX (CUDA)	Jetson GPU — cross-platform	11.4ms · 87fps
TFLite INT8	CPU-only ARM, bandwidth constrained	19.2ms · 52fps
ONNX (CPU)	Mixed fleet, single file	48.2ms · 20fps

⚠

TensorRT engines are device-specific. A TRT engine compiled for Jetson Orin will not run on Jetson Nano. Register hardware-specific variants using mlops push --variant and the agent selects the correct file automatically.

Compliance & audit logs

Every deployment event, drift alert, and rollback is recorded in an immutable audit log. Required for FDA SaMD (21 CFR Part 11), ISO 13485, and CE MDR Article 10.

BASH — Export audit log

# Export audit log for a device (JSON or CSV)
mlops audit export   --device jetson-prod-01   --from 2025-01-01   --to 2025-07-15   --format csv   --output audit-q1q2.csv

Enterprise customers receive IQ/OQ/PQ documentation templates and a dedicated ML infrastructure engineer for FDA/CE submission support. Email hello@mlops.dev.

REST API

Base URL: https://api.mlops.dev/v1. All requests require Authorization: Bearer <api_key>.

Endpoint	Description
GET /devices	List all registered devices
GET /devices/:id	Device status, model version, drift score
GET /models	List model registry
POST /models	Register a new model version
POST /deployments	Create a deployment
GET /deployments/:id	Deployment status and stage progress
POST /deployments/:id/rollback	Trigger rollback
GET /drift	Drift scores across fleet
GET /audit	Audit log (paginated)
GET /health	API health check

ℹ

Full API reference with request/response schemas is available to waitlist members. Join the waitlist →