Selected work.
A short list of what we are willing to talk about publicly. The full reel goes out under NDA on a first call. Most clients prefer it that way and we respect it. Our own office liaison build sits at the top because it is the one project we can write about in the open, in full, with the metrics intact.
An office liaison that actually does the job, documented in public.
The problem
An AI integration shop that does not deploy its own robots reads like a marketing site. We bought a Unitree Go2 EDU+ in April 2026 and gave it a real job: greet office visitors, escort them to meeting rooms, answer basic questions on its own. Same SLA we would commit to for a paying customer.
Constraints
The Castle Rock office has a frosted glass partition, polished concrete floors, and a long corridor with high ambient noise. The robot had to navigate all of that safely with guests in the room, hit a sub-four-second conversational response time, and run with zero cloud dependency when the WiFi flakes (it does).
The build
Two weeks from unbox to live. ROS 2 Humble on Ubuntu 22.04 with a micro-ROS bridge over DDS. Whisper-small for ASR onboard, Piper for TTS onboard, a quantized Llama 3 8B on the Orin for the conversational layer with a Claude 4 fallback when the question deserves it. Custom Python orchestrator on top, RTAB-Map for SLAM with a hand-rolled glass-mask filter for the partition. Prometheus and Grafana for telemetry, Sentry for alerts.
What we would do differently
We would mount the rear-facing depth camera higher. The default placement on the EDU+ misses anyone under five feet at close range, which made child interactions awkward for the first three days. We would also skip the cloud-fallback architecture entirely for v2. Local-only is faster, more reliable, and reads better in a security review. The Orin can handle it.
283 conversations logged last 30 days
67 visitors escorted first 14 days deployed
3.8s mean first response down from 6.4s cloud round trip
94% task completion rate against the eval suite
0 emergency stops since deployment
0.34m SLAM lock precision inside the lobby
SPEC SHEET The actual stack
- Robot
- Unitree Go2 EDU+
- Compute
- NVIDIA Jetson Orin (onboard)
- OS
- Ubuntu 22.04 + ROS 2 Humble
- ASR
- Whisper-small.en, onboard
- LLM
- Llama 3 8B Q4 (local), Claude 4 fallback
- TTS
- Piper, onboard
- VLM
- Moondream 2
- SLAM
- RTAB-Map + custom glass-mask layer
- Bridge
- micro-ROS over DDS
- Telemetry
- Prometheus + Grafana, Sentry alerts
Four engagements we cannot name yet.
Each of these is a real client under NDA. We will rewrite this section when the contracts allow. Until then, here is the rough shape.
Luxury auto showroom walk-around at scale
UNDER NDAData-center scheduled walk patrol
UNDER NDANRF booth humanoid with branded agent
UNDER NDA Get started
Want to see the full reel?
A first call gets you the NDA. The NDA gets you the rest of the work, named clients, real numbers, real timelines.