We build AI systems that actually work.
CairoLabs.ai is our engineering practice for applied AI work — autonomous development pipelines, local-model infrastructure on real hardware, and practical tools, built, documented, and shipped in public.
Humans in the loop, always. Hallucinations, never.
AI agents plan and build; we review and decide. Every pipeline we ship escalates consequential choices back to a person before they execute. Not because the models aren't good — because that's the right architecture. The goal is leverage, not abdication.
Our products and projects — standalone tools, plus the infrastructure they run on. Built and running on our own hardware, and available to run on yours.
Project Lumbergh
A multi-tier autonomous dev pipeline, staffed by AI and themed — of course — after Office Space. Humans sit at the C-Level; local VP agents (aka Bill Lumbergh) plan and decompose the work; PMs dispatch and manage parallel code builders (aka Peter); Milton runs maintenance, housekeeping and bug-squashing from the basement; and The Bobs — a panel of outside models — review and approve high-stakes changes.
The models talk to each other over custom NATS messaging and escalate to humans through Slack. Local models run on NVIDIA DGX Spark workstations that mirror code to GitHub and feed the TPS Reports dashboard, which monitors builds and raises alerts. Every escalation path ends at a human signature. Sparky came online in November 2025; Lumbergh has been filing and building real work since March 2026, self-contained and self-fixing:
- 1,750 intakes filed
- 7,700 tasks executed
- 4,700 autonomous AI coding sessions
- 4,500 automated QA reviews
- 12,000 git commits
- 630,000 lines of code
- 820 bugs filed, 640 shipped
The Bobs
A subtle secret to using AI is that each improvement in the prompt you provide generates exponentially better outcomes. It's the classic "Garbage In, Garbage Out" scenario, amplified. The Bobs are a panel of independent AI models — both frontier and local — that adversarially and independently review high-stakes specs and code, surfacing disagreements and weak spots in the design before anything gets built.
A synthesizer AI then compares the panel's findings, fact-checks the discrepancies and low-confidence results, and produces a research-grade design spec that eliminates bugs and hallucinations before they reach code. The Bobs in Office Space were outside consultants who reviewed the plan and reported back — the perfect foil for Project Lumbergh, and the gate every major intake passes through.
Photo Tagger
Keywording a photo library by hand isn't a project, it's a sentence — and it only gets worse as the catalog grows. With over 200,000 images of his own and forty years behind the camera, Mike built this one out of necessity.
Photo Tagger runs a small purpose-built AI model on your own machine — no cloud, nothing uploaded — and auto-tags Adobe Lightroom Classic libraries so they are actually searchable. It watches folders to tag new imports, learns and reuses your own keyword vocabulary, writes your copyright and creator metadata, and can back up your tags to the cloud while the photos never leave your machine. Start free, add image packs as needed, or go unlimited with Personal or Pro.
Other things we've built
- Frame TV Image Uploader A desktop tool that prepares and one-click uploads your own photographs to a Samsung Frame TV's Art Mode, handling Samsung's finicky resolution and colour requirements for you, with matte styles you can preview first. Open source. On GitHub →
- The AI Weekly A curated digest of the week's most useful AI news, tips and lessons, generated in-house on our own pipeline from what our three private Cairo user groups are actually sharing. Read it on Cairoglyphics →
- DGX Spark Lab The in-house NVIDIA DGX Spark workstations where we build and test with production-grade local models — plus the internal tool that benchmarks each new model against our real workflows, so upgrades are decided on evidence rather than hype. More in our writing →
- ONNX Runtime — ARM64 Blackwell PTX The first ONNX Runtime + TensorRT stack we know of for Grace Blackwell ARM64, built with a PTX-JIT process to work around missing driver support. Archived when upstream issues beyond our control made it moot — shared anyway, honest dead-ends included. onnxruntime-arm64-blackwell-ptx →
Our writing lives at Cairoglyphics.ai.
Build logs, essays, the AI Weekly newsletter, and podcasts to come — all at Cairoglyphics.ai, the shared hub for CairoLabs.ai, CairoConsulting.ai, and Cairo Networks LLC. The link below opens Cairoglyphics filtered to Cairo Labs; clear the filter there to read everything we publish.
The Cairo Companies currently consist of only two people, by design — Mike and Helen, co-founders. Mike runs tech, vision, and roadmap; Helen builds AI tech to help run our business operations. We personally oversee all AI development to ensure we exceed our customers' expectations.
Mike has spent thirty years in IT. After joining Aruba Networks in 2007, he architected and managed enterprise network deliveries for over 15 years, then joined a network-as-a-service startup where he helped build out the products and the Professional Services offering — writing pricing and scoping and structuring the consulting model from scratch. He watched AI rewrite the network engineer's job from the inside, and decided to build with it rather than around it.
Helen spent over a decade in business operations. At a scaling SaaS company she ran HR, investor relations, and legal due diligence, owned the contract and equity software stacks, and managed vendor relationships. She has rolled out enough cloud software at growing companies to tell a real productivity tool from a marketing campaign.
Mike and Helen's career experience has allowed us to rapidly build three related AI businesses. But when a client project needs hands beyond our scope, we bring in skilled, vetted specialists from a Rolodex we've spent forty years building — no work farmed out to strangers. Every person we put on something is one we'd hire ourselves.