Private AI
by Design
Intelligence belongs where your data, code, and decisions already live—not across a public API boundary. SMILE unifies Local AI at the developer and analyst desk with Edge AI in production gateways, branches, and self-hosted clusters, running natively on the JVM and GraalVM without cloud lock-in or Python sprawl.
Two surfaces, one architecture
Local AI at the desk. Edge AI in production.
Private AI is not a single deployment box. It starts where builders explore data, write code, and run autonomous workflows—and extends to the production servers and edge nodes where models score live traffic:
Human-in-the-loop & autonomous bots
Local AI — SMILE Agent Studio →
For indie hackers, AI-native startups, and regulated data science or JVM engineering teams. Run nine coordinated specialist agents, 24/7 autonomous Bots (powered by Agent Hub), and interactive notebooks directly against the files, databases, and git repositories on your machine.
Machine-to-machine & production runtime
Edge AI — Serve & Core Engine →
For payment networks, bank branches, telematics fleets, ISVs, and air-gapped sovereign systems. Embed SMILE algorithms directly inside your JVM application or deploy a single-process Quarkus inference server—scaling from an 80 MiB GraalVM native binary (0.2 GB image, 0.4 s boot) to multi-GPU LLM clusters.
The Private AI Vision
Why intelligence is moving back inside your boundary
- Data gravity and sovereignty are non-negotiable Cardholder transactions, patient records, proprietary trading signals, and unreleased source trees cannot be streamed to multi-tenant SaaS wrappers. Whether you are a two-person startup protecting your core IP or a global institution under PCI-DSS, HIPAA, or sovereign air-gap mandates, bringing the model and the agent to the data is safer, faster, and cheaper than shipping the data to the cloud.
- Real-time decisions cannot afford a cloud round-trip Payment authorization, fraud scoring, industrial telemetry, and tight multi-step agent loops run on millisecond budgets. Executing classical ML, ONNX graphs, and local LLMs in-process—with persistent draft prefixes, radix KV caching, and local TreeSHAP / KernelSHAP attribution—eliminates network jitter and per-token API metered billing.
- No Python sidecar tax in production Mission-critical enterprise backends and desktop tools run on the JVM, while AI tooling has been fragmented across fragile Python virtual environments and multi-gigabyte container stacks. SMILE unifies classical ML, deep learning, ONNX Runtime, and LLM inference in a clean Java 25 and GraalVM native architecture—while still giving Python engineers a first-class notebook and agent home in Agent Studio.
-
Explainable and auditable where the decision happens
Private AI is not just about keeping bytes behind a firewall; it is about proving why an automated decision was made. SMILE computes exact TreeSHAP, LinearSHAP, and KernelSHAP attributions in-process alongside every prediction and keeps structured agent memory (
.smile/memory/) in plain markdown on your own disk. -
100% ownership of your models, artifacts, and workflows
Every notebook, codebase change,
SKILL.mdworkflow, and exported.smlor.onnxmodel artifact belongs entirely to you. With SMILE's commercial licensing tiers, you can run in production or embed inside shipped products with zero copyleft disclosure and zero per-prediction royalties.
End-to-end closed loop
How Local AI and Edge AI work together
SMILE is architected so that what you build at your desk runs unmodified in production—without rewriting pipelines from Python into C++ or Java:
| Stage | Where It Runs | SMILE Component | What Happens |
|---|---|---|---|
| 1. Explore, Architect & Train | Local AI (Developer / Data Scientist Workstation) |
Agent Studio | Profile local CSV/Parquet/DuckDB data, engineer features in Java/Python/Scala/Kotlin notebooks, collaborate with Clair, James, Guido, and Ada, and export verified .sml and .onnx models. |
| 2. Automate Operations 24/7 | Local AI (Workstation or Private Team Server) |
Autonomous Bots (smile hub) |
Run scheduled and event-driven bot routines in bots.json (cron timers, emails, calendar events, webhooks) across engineering, data validation, product, growth, and sales. |
| 3. Embed In-Process | Edge & Backend AI (Inside Your Application Binary) |
Core Engine | Call SMILE algorithms, local TreeSHAP/KernelSHAP, ONNX, and LLM inference directly inside your Java, Scala, or Kotlin application with zero external service hops. |
| 4. Serve Across Nodes & Fleets | Edge & Cluster AI (Gateways, Branches, Private Cloud) |
Serve Engine | Deploy ghcr.io/haifengl/smile-serve (latest-micro 80 MiB native binary up to latest-gpu CUDA clusters) with hot-swappable models, OpenAI-compatible chat, and Prometheus telemetry. |
Start local, scale to the edge
Pick where you want to start
Evaluate Agent Studio and the open-source engines on your machine today, request a 30-day Edge AI pilot on your own nodes, or review our upfront commercial licensing for Production Runtime, Perpetual OEM, and Agent Studio teams.