We publish open research at the intersection of AI and spatial reasoning — contributing tools, frameworks, and ideas that help anyone build smarter infrastructure for the real world.
Most AI research focuses on language and vision. We focus on the third modality: spatial reasoning over the physical world. Our work sits at the intersection of large language models, mathematical optimization, and geospatial systems.
We believe the most impactful AI research will be work that helps everyone — not just us — build systems that understand physical constraints, plan infrastructure autonomously, and make critical spatial decisions with precision. We open-source our findings so the entire community can build on them.
Encoding geographic context — distances, topologies, physical constraints — as structured LLM input. Enabling any AI system to reason about the real world, not just text.
A single agent design that adapts to any infrastructure domain through schema definitions. Our approach shows that domain expertise doesn't require domain-specific code.
Combining MIP, constraint programming, VRP, graph algorithms, simulation, and heuristics — published so others can apply these techniques to their own infrastructure challenges.
Open research on spatial AI, autonomous design, and optimization — published to advance the field and invite collaboration.
A. Aviv, Nexma Research
We introduce a framework for encoding geographic context — distances, topologies, physical constraints — as structured input to large language models. Our approach enables LLMs to reason about spatial relationships with the precision required for infrastructure engineering, achieving constraint-compliant designs in 94% of test cases across four infrastructure domains.
A. Aviv, Nexma Research
We present Jax, an AI agent that autonomously designs physical infrastructure using domain-agnostic ontology schemas. By decoupling domain knowledge from agent architecture, a single agent serves telecom, water, electric, and logistics domains. We demonstrate 36x speed improvement over manual design with equivalent constraint compliance.
A. Aviv, Nexma Research
We describe a hybrid optimization approach combining MIP formulations with geography-aware heuristics for FTTH network design. The solver respects TIA-598 color coding, GPON cascade depth limits, and optical budget constraints while minimizing total fiber length. Results show 23% cost reduction vs. commercial planning tools across 12 real-world deployments.
A. Aviv, Nexma Research
We propose the GeoCodebase — an in-memory virtual file system where spatial data, schemas, and configuration coexist as a unified text representation. AI agents interact through file primitives (read, write, grep, glob) mirroring developer workflows. This design eliminates domain-specific tool proliferation and enables schema-driven UI generation from a single source of truth.
Technical deep-dives
Over two years of building AI agents that design fiber networks on real maps, we learned something most LLM-agent literature ignores: the physical world doesn't tolerate hallucination. This post breaks down the five fundamental constraints and the design principles that survived production.
Traditional FTTH planning requires weeks of manual work across GIS, CAD, and spreadsheet tools. Nexma's agentic approach compresses this into minutes — a single AI agent reasons about topology, optical budgets, and real-world street geometry to produce deployment-ready designs.
Splice documentation is one of the most error-prone steps in fiber deployment. We built an AI-driven splice engine that generates fully compliant TIA-598 diagrams, tracks every fiber from OLT to ONT, and validates optical loss budgets before a single cable is pulled.
Mixed-Integer Programming finds mathematically optimal solutions — but it doesn't know about building footprints, street access, or pole-mounted closures. We combine MIP with geography-aware heuristics to get the best of both: optimality where it matters, practicality everywhere else.
Open collaboration
We're looking for researchers and engineers who want to teach AI to understand the physical world. Remote-first. Research-driven.