Login

The engine behind Jax

Meet the Math Engine

The engine that translates natural language into mathematically optimal spatial solutions — and commits nothing it has not verified against the real-world constraints of each domain.

01
The problem

Mathematical programming is fundamental across domains, yet remains a skill-intensive bottleneck

Mathematical optimization plays a critical role across many business sectors, from supply-chain management to energy systems to logistics planning, where effective decision-making relies on solving highly complex optimization problems.

While practitioners can usually describe these problems in natural language, translating them into precise mathematical formulations that optimization solvers can process remains a skill-intensive bottleneck. Crafting a correct formulation requires precise definition of decision variables, objectives, and constraints — a skill that typically takes years of specialized training in operations research to develop.

Our approach

Math Engine automates this task — translating natural language into executable optimization models, dispatching to the right solver, and self-correcting against domain constraints. No operations research expertise required.

02
Math Engine

General-purpose AI describes infrastructure fluently, yet returns designs that fail on inspection. Math Engine is purpose-built to produce designs that are verified buildable.

It separates two concerns that are usually conflated. Generality comes from an ontology — a typed model of a domain's entities, relationships, and constraints. The same agent engineers fiber, water, power, or logistics by reading the schema, not by being rebuilt for each.

Feasibility comes from verifiable optimization. Every decision is formulated as a classical optimization problem, solved deterministically, and validated against the domain's physical constraints — nothing is committed until it has passed.

The result is a plan that is buildable, not merely plausible — produced in a single session, and checkable line by line against the standards each domain demands.

03
Performance

Lower build cost

Designs minimize total installed cost — equipment, cable, and civil works — by solving for the cheapest arrangement that violates no constraint.

Faster design cycles

Network designs that take human engineers days to weeks are produced in a single session, with full constraint verification.

Many domains, one agent

Fiber, water, power, wireless, logistics — the same agent engineers each by reading its ontology. Change the schema, not the code.

Architectural design goals, not benchmarked results. Measured solver runs across domains are reported in the research paper below.

Fig 1.A — Math Engine

0%

[FIG 4.A.1] Formalize design intent

Natural language design requirements — coverage targets, equipment constraints, budget limits — are parsed into a formal optimization specification: decision variables, constraints, and objective functions that mathematical solvers can process.

[FIG 4.A.2] Encode spatial context

The physical environment — street geometry, building footprints, existing infrastructure, terrain — is structured into a machine-readable spatial representation that the AI reads and reasons over through deterministic queries.

[FIG 4.A.3] Classify problem structure

The formalized specification is mapped to canonical optimization classes — facility location, network flow, vehicle routing, scheduling — ensuring the right solver family handles each subproblem.

[FIG 4.A.4] Generate optimal design

The selected solver produces a mathematically optimal solution — equipment placement, cable routing, resource allocation — as structured, auditable operations on the spatial representation.

[FIG 4.A.5] Verify against constraints

Every generated design element is validated against engineering standards: capacity thresholds, maximum distances, physical laws, regulatory limits, and budget constraints. Violations are detected and repaired automatically.

[FIG 4.A.6] Converge on minimum cost

A mixed-integer programming solver evaluates the design against alternative configurations, converging on the solution that minimizes total deployment cost while satisfying every validated constraint.

04
The loop

One agent,
one verifiable loop

Math Engine is not a suite of separate models. It is a single agent running one loop for every decision it commits — recognize the problem, formulate it, solve it, then validate and self-correct before anything is trusted.

[1]

Recognize

From the intent and the domain's ontology, the agent identifies which class of optimization the goal implies — facility location, routing, scheduling, network flow, or steady-state simulation. The same recognition runs for every domain; what changes is the ontology it reads, not the agent.

Specification

1.1Classify problem structure from natural language
1.2Map intent to a canonical optimization family
1.3Match on mathematical structure, not keywords
1.4One unmodified agent across every domain
[2]

Formulate

The agent constructs a concrete optimization instance — decision variables, objective, and constraints — populated entirely from the ontology and expressed in a uniform problem contract. Class-specific guidance steers it away from the formulation errors characteristic of each family.

Specification

2.1Build variables, objective, and constraints from the typed world model
2.2Grounded in mathematical programming, not approximation
2.3Class-specific formulation guidance per family
2.4No operations-research expertise required of the user
[3]

Solve

The instance is dispatched to a deterministic solver that returns a feasible — and, where an objective exists, optimal — result, with its optimality gap. A small family of classical solvers covers the full spectrum, and engine selection is automatic from problem structure.

Specification

3.1Deterministic solvers across the optimization families
3.2Automatic engine selection from problem structure
3.3Optimal or near-optimal, with a reported gap
3.4Correctness independent of where it solves
[4]

Validate & self-correct

Every solved result is checked against the domain's full constraint set — including physical checks the formulation may not encode. Nothing is committed until it passes; a failed check returns named diagnostics that drive a bounded re-formulation. The design that ships is the first one that passed, not the first one produced.

Specification

4.1Validate against engineering standards and physical laws
4.2Commit nothing that has not been verified
4.3Failed checks drive bounded self-correction
4.4Honest escalation when no feasible design exists
06
Leadership

Built under the leadership
of a world-class spatial AI expert

Building technology of this caliber requires deep expertise in operations research, mathematical optimization, and spatial intelligence — combined with the ability to architect an entire AI agent platform from the ground up.

AI leadership

Ari Aviv

Founder & CEO

07
Security

Engineering-first approach
to data privacy and security

We have invested heavily in data privacy and security. All infrastructure designs are generated within your project environment.

  • Geospatial data is encrypted in transit and at rest
  • No customer data is used for model training
  • SOC 2 compliance roadmap in progress
  • Project-level isolation — your data is never shared across workspaces

Learn more

Get purpose-built AI
for industry-leading
physical world intelligence