01What we build

A small team of senior engineers. We go deep on three things.

  1. Applied AI

    LLM apps, agents, vision

    LLM agents, retrieval, and vision-language models in production. We build the evals and guardrails that keep them working outside the demo.

    • LLM agents
    • RAG
    • VLMs
    • Evals
    • Fine-tuning
    • Real-time inference

    Shipped: The Rivals, VCC SIM, IEEE Competitions

  2. Robotics & Embodied AI

    Real robots, real streets

    Autonomous navigation, dexterous manipulation, teleoperation, SLAM, sim-to-real, and fleet management. We co-author the open-source Earth Rovers SDK that real-world AI models like GeNIE (winner, Earth Rover Challenge 2025) run on.

    • ROS 2
    • SLAM
    • Manipulation
    • VLA models
    • Imitation learning
    • Sim-to-real
    • Foundation policies

    Shipped: Earth Rovers, UFB, Telearms

  3. Proteus, our maritime collision-avoidance overlay fusing AIS, radar and ENC chart data.

    Connected & Edge Systems

    Software that lives on hardware

    On-device inference, sensor streaming, BLE and MQTT telemetry. Cloud, edge, and hardware held to one engineering bar.

    • BLE
    • MQTT
    • WebRTC
    • AWS IoT Core
    • NMEA 2000

    Shipped: Tocaro Blue, Proteus

02At the frontier of physical AI

We publish, organize, and open-source the work behind physical AI.

  • Earth Rover Challenge organizers and participants at IROS 2024 in Abu Dhabi.
    IEEE Robotics & Automation Magazine

    Co-author, Lessons Learned From the Earth Rover Challenge

    Peer-reviewed paper on an AI vs. human navigation competition across 14 cities.

    Read the paper
  • Open source

    Earth Rovers SDK

    We co-author it. Labs at UC Berkeley, NUS and UT Austin deploy real-world robot policies on it.

    View on GitHub
  • IROS 2026, Pittsburgh

    Co-organizing three robotics challenges

    Autonomous navigation, dexterous manipulation, and humanoid assembly.

  • Open data

    9,000+ hours of real-world robot teleoperation

    FrodoBots-2K and Berkeley-FrodoBots-7K, the largest open sidewalk-robot dataset.

    Browse on Hugging Face

Competitions and datasets are run by the FrodoBots-led coalition; our founder is a co-organizer and IEEE co-author.

03Selected work

04How we engage

Three ways to work together.

  1. Scoping sprint

    1 to 2 weeks, fixed

    An architecture, a costed plan, and a working prototype. Yours to keep, even if we stop there.

  2. Build

    3 to 6 months

    We ship to production. Weekly demos, code you own, your team in the loop.

  3. Long-term partnership

    Monthly retainer

    Maintenance, scaling, new features, on-call. Most clients renew past year one.

05Engineering led by people, not playbooks

  • Santiago Pravisani, CEO and co-founder

    Santiago Pravisani

    CEO and co-founder

    Core engineer at FrodoBots. Co-organizer of three IROS 2026 challenges. Contributor to the Earth Rovers SDK.

  • Esteban Fuhrmann, CTO and co-founder

    Esteban Fuhrmann

    CTO and co-founder

    Leads architecture and delivery across every SANA project.

06Tell us your project