Projects

Systems that had to work
where nobody was watching.

A selection of programs I architected or led — each one with the illustration of what was actually delivered, and links out to the platforms, products and organizations behind them.

“The Deployment Stack for Physical AI” — what Cayley builds, and why the layer under the models is the one that decides whether a fleet ever reaches the field.

Flagship Cayley AI Present

The physical AI deployment stack

Co-founded to build the infrastructure required to scale autonomous robotics across heavy industry and defense. Models keep getting better faster than the systems that can actually field them — Cayley works on that second half.

  • Architecting resilient infrastructure and vehicle autonomy for reliable telemetry and control in unstructured environments
  • Driving edge orchestration across distributed, intermittently-connected assets
  • Standardizing embedded hardware ecosystems so deployment stops being bespoke per vehicle
  • Integrating complex sensor suites to enable seamless multi-asset coordination
  • Targeting the cases where traditional solutions fail outright or become cost-prohibitive
The five-layer robotic deployment stack, bottom to top: infrastructure, operation, machine and perception, all reusable across deployments, with a custom application layer on top.
Four reusable layers carry every deployment; only the application on top is built per customer.

References

Autonomous Caterpillar dozers and a grader working a large earthmoving site, seen from above.
Heavy machinery working a live site — the fleet a single operator has to be able to hold.

HavocAI (Teleo)2024 — 2025

Supervised autonomy for heavy vehicle fleets

Industrial sites run on heavy machinery and the people who sit in it. Supervised autonomy only changes that economics if one operator can hold several machines at once — which turns out to be a communications problem long before it is an autonomy problem.

  • Architected ultra-low-latency, high-bandwidth communication infrastructure spanning expansive industrial sites
  • Engineered and optimized high-performance video transmission pipelines
  • Enabled a 1:N operator-to-machine ratio, significantly increasing operational throughput
  • Bridged sales, deployments and operations so infrastructure requirements matched what the field actually needed

References

A V-BAT uncrewed aircraft hovering above a ship's flight deck at sea before landing.
V-BAT recovering to a deck at sea — the landing case that precision localization exists to make routine.

Shield AI2022 — 2024

Autonomy integration & avionics on the V-BAT

Getting an autonomy suite onto a flying platform is a systems problem: the software has to land on hardware that is reliable, updatable, standardized across a fleet, and provably ready to fly before anyone launches it.

  • Autonomous systems integration — led integration of the Hivemind AI autonomy suite onto the V-BAT platform, enabling mission-level intelligence
  • Avionics modernization — re-architected legacy avionics and ported firmware to a high-performance distributed microcontroller platform
  • Organizational standardization — established the NVIDIA Jetson ecosystem as the org-wide edge-compute standard, streamlining AI model deployment across the fleet
  • Vehicle health & safety — engineered VHM systems and automated pre/post-flight evaluation for a 100% operational readiness envelope
  • Precision localization — developed novel short-to-medium range positioning for high-assurance landing in complex or contested environments
  • Mission-specific AI evaluation — analyzed vision and AI platforms for high-endurance, wide-area search

References

Ultrasonic thickness map of a tank floor plate rendered as a colour heatmap — greens and yellows across the bulk of the plate with blue thin-wall regions and red weld seams at the edges.
Delivered thickness map — a single floor plate, reported at 1 mm × 1 mm grid resolution.
Colour encodes remaining wall thickness; blue regions are the corrosion the survey exists to find.

Tankbots · Advanced Robotics Group 2020 — 2022

In-service autonomous tank inspection

A first-of-its-kind autonomous underwater vehicle built to inspect storage tanks without taking them out of service — no cleaning, no confined-space entry, no man-hours inside a hazardous vessel. The vehicle runs 24-hour endurance missions in an environment with zero connectivity: no GPS, no RF, no operator link once it goes in.

  • Architected the AUV for long-endurance missions in fully signal-denied conditions
  • Developed a proprietary underwater visual odometry system for high-precision navigation and localization
  • Engineered full-stack autonomy pipelines capable of complex decision-making with no human in the loop
  • Turned raw acoustic returns into an inspection-grade deliverable an integrity engineer can act on

Delivered on a single survey

  • 244,792,225thickness readings
  • 32,099,270qualified measurements
  • 14,750,144reported 1 mm × 1 mm grid cells
  • 99.7%surface coverage

Roughly one qualified measurement for every eight raw returns — the gap between those two numbers is the qualification pipeline, and it is where inspection data becomes evidence instead of noise.

References

Wireframe visualization of a rig floor with LiDAR-derived geometry, machinery and a person highlighted in green, with hazard zones shaded red.
Live LiDAR reconstruction of a working deck — people in green, hazard zones in red, computed on the machine.

Microsoft · The Marsden Group2019 — 2020

Embedded AI, perception & network optimization

Making real-time perception fit inside the power and latency budget of a machine in the field — then making sure the data could actually get off it. The work shipped as VisionIQ, a LiDAR-and-vision safety system for offshore rig floors.

  • Optimized CNN inference on NVIDIA Jetson platforms using Tensor Cores: 4× faster at lower power
  • Built high-performance LiDAR processing pipelines with a 500× improvement in execution time
  • Achieved a 20× reduction in network latency through kernel-level and TCP stack tuning
  • Designed adaptive camera algorithms for real-time exposure and focus control in variable lighting
  • Built self-discovering wireless mesh protocols with dynamic load balancing
  • Managed end-to-end hardware/software integration: sourcing, thermal management, physical optimization

References

Aerial view of a stimulation site: rows of pump units, blenders and manifold trailers laid out around a wellhead.
A stimulation spread from above — the high-pressure iron the platform actuates, and the red zone nobody should be standing in.

Schlumberger · SLIIC2016 — 2019

ValveCommander

A cloud-controlled, diesel-engine-powered hydraulic and electric actuation system for high-risk, high-pressure operations. I managed its execution across the entire product lifecycle — inception and architecture through engineering, manufacturing, field trials and commercialization.

  • Owned the product end to end, from first architecture to commercial release
  • Awarded a top-5 innovation project across the entire company
  • Has been generating revenue for over a decade
  • Also built a custom testing system for characterizing nuclear detector photomultipliers
  • Created a high-precision software suite to quantify radiation detector accuracy for prototype validation

References

Diagram of a downhole perforation event captured over optical fibre, processed by an FPGA, buffered on an ARM microcontroller and transmitted over TCP/IP to a receiver.
Fibre to FPGA to MCU to the network — a capture chain measured in microseconds.

Schlumberger · Coiled Tubing2014 — 2015

High-frequency downhole acquisition

A device to capture high-frequency data from underground explosive perforations — transmitted over optical fibre, processed on FPGAs, handed to an ARM microcontroller and streamed over TCP/IP to a receiver at surface.

  • Designed the full acquisition chain across optics, programmable logic, firmware and networking
  • Then took it to the field: rigless well intervention with a team of five in the remote deserts of Saudi Arabia
  • Hands-on with fluid pumps, diesel engines, industrial hydraulics, liquid nitrogen and concentrated acid — the operational reality the software has to respect

References

Next

Working on something in this shape?

If your program looks like one of the above — a vehicle that has to think for itself, a fleet that has to be operable, or a sensor stack that has to become a deliverable — I take on a small number of consulting engagements each year.

Consulting & advisory marcin@rutkow.ski