Learn · Three Concepts · Five Minutes

What makes AI math-grounded.

Three ideas, in order. No calculus required, but if you know some it will read extra well.

Closed-form beats learned, when you can get it.

A neural network is a function. So is a Lyapunov controller. So is an RSS envelope. So is a Kalman filter. Some functions you learn from data. Others you derive from principles.

When a physical-AI primitive has a closed-form derivation, use it. You get a proof, a bound, and a receipt. You don't get emergent behavior where you needed a guarantee.

RSS safety envelope — closed form

d_min = v·τ + (v² / (2·a_min))
# ego speed v, reaction time τ, deceleration floor a_min
# no training data. no weights. just physics.

Lives in mgai-auto. Returns a Metered<f64> with the value and its picojoule cost attached.

Every call is metered in picojoules.

In MGAI, a computation doesn't return a value. It returns a value and its energy cost. Cost is computed at the call site from a cost model tied to a specific silicon target (e.g., auto-grade 7 nm).

You don't guess what your pipeline will cost on a battery-powered drone. You know. Before you ship it, not after.

A metered call

let m = safe_longitudinal_distance_metered(
    v_ego, v_lead,
    &RssParams::highway_defaults(),
    fuzz,
    &CostModel::AUTO_GRADE_7NM,
);

assert_eq!(m.energy_pj.0, 75);  // 75 picojoules per call
let envelope = m.value;         // meters

Across a 400-step integrated run: 37,200 pJ cumulative. Times, call counts, and per-stage breakdowns live in mgai-telemetry.

Primitives compose because they share types.

MGAI is one Cargo workspace. Every crate depends on mgai-core. Extension points like PipelineStepExtensions are serde-backward-compatible Option<T> slots, so any vertical can add a field without breaking anything downstream.

Safety-feeds-auto-feeds-fleet isn't an architecture diagram. It's a type signature.

One struct, three verticals

pub struct PipelineStepExtensions {
    // populated by mgai-auto (RSS envelope, per-frame)
    pub av_safety:       Option<AvSafetyExtension>,

    // populated by mgai-vla (DAM dual-head inference)
    pub vla_inference:   Option<VlaInferenceExtension>,

    // populated by mgai-digital-twin (divergence report)
    pub twin_divergence: Option<TwinDivergenceExtension>,
}

Same trajectory, three independent verticals, zero glue. That's the product.

That's it.

Closed-form. Metered. Composed.

Watch the film Open the playground →

Or go straight to the code

git clone https://github.com/openie-dev/mgai
cd mgai && cargo test