We build the software your machines run on — and we keep it up.
Uptime Systems AI is an engineering team for companies that ship physical products. Custom software and applied AI, delivered by the sprint, then operated with real SLOs, real on-call, and a compliance file that survives an audit.
The same telemetry your team would see
From a register on a PLC to a decision on a screen — every hop named
This is the shape of nine delivered systems, with the protocol at each boundary, the retention at each store and the p95 budget for each hop. Zones follow IEC 62443; levels follow ISA-95. Nothing here is aspirational — it is the drawing we hand to your plant IT on day three.
Three ways to use us. Most clients use two.
No discovery theatre, no forty-slide readouts. You get engineers who have shipped this class of system before, working in your repo from week one.
{{ md.body }}
The stack, end to end
{{ c.body }}
Systems in production, with the numbers
{{ cs.body }}
Two weeks to a running slice. Then every two weeks after that.
We do not sell a discovery phase. Sprint one ends with something deployed to your staging environment and a demo you can hand to an operator.
We run a datacenter in Ashburn, Virginia. Your workload can live in it.
Not a reseller badge on someone else's cloud. A small facility we own and staff in Loudoun County, forty minutes from our Virginia office, so a regulated or export-sensitive workload can run on US soil, on hardware with a serial number we can point at, operated by US persons.
Notes from the work
{{ a.dek }}
Tell us the system and the constraint. We will tell you the shortest path to it running.
Forty-five minutes with the engineer who would lead the work. You leave with an architecture sketch, a sprint-one scope, and a straight answer on whether we are the right team.
MIN ENGAGEMENT: 4 SPRINTS
Senior engineering capacity, rented by the sprint.
You have a roadmap and not enough engineers who have built this before. We drop a pod into your repo, your standups and your on-call rotation, and it starts landing merges in week one.
Pod shapes
{{ pd.body }}
Where a pod plugs in
SCROLL →What we will not do
Being clear about this early saves both of us a quarter. If any of these is what you need, we will say so on the first call and point you somewhere better.
Tell us the stack and the deadline. We will tell you honestly what we can staff and when.
The system that replaces the spreadsheet two people understand.
Every hardware company runs on a handful of workbooks, shared drives and one heroic engineer's macros. We turn those into software with an owner, a schema and a test suite.
{{ b.body }}
Default stack, and when we deviate
We have opinions so you do not have to spend a sprint on them. They are defaults, not religion — if your team already runs .NET and Azure, we work there instead.
A build, in eight weeks
AI that is measured before it is announced.
We build the unglamorous kind: extraction that reads your supplier certificates, detection that catches a drifting tool, retrieval that answers from your own manuals. Every one ships with an evaluation harness your team can run without us.
Where it earns its keep
{{ u.body }}
The evaluation harness, in plain terms
SCROLL →The harness lives in your repository and runs in your CI. If we stop working together tomorrow, you can still tell whether the model got worse.
Honest limits
The fastest way to waste a quarter on AI is to point it at a problem it cannot hold. Here is where we push back.
Off the floor, into one schema, without a nightly CSV in sight.
Most industrial data problems are not modelling problems. They are five protocols, three clocks and a serial number that means something different in each system.
Protocols we meet you on
Data contracts, not data lakes
Every feed gets a contract: owner, schema, freshness target and what happens when it breaks. A dashboard nobody trusts is worse than no dashboard.
The name in the company is uptime. This is the part we are named after.
Software that touches physical equipment fails in ways web software does not: a stale reading is worse than no reading, and a retry can move a real actuator twice. We engineer for that, then carry the pager.
A real SLO sheet
CLIENT DETAILS REDACTEDWhat an incident looks like from your side
{{ s.body }}
On-call tiers
Monthly reliability review
Starting with a reliability audit: two weeks, fixed fee, and a ranked register of what will page you next quarter.
Compliance as pipeline output, not a quarterly fire drill.
Obligations arrive as documents. They are satisfied by engineering. We map each requirement to a control that runs in CI and an artefact with a version number, so the evidence pack is a build output rather than a two-week scramble.
Obligation → control → evidence
SCROLL →Positions summarised as of September 2026 and not legal advice. Dates for the Cyber Resilience Act and the AI Act are those published in the Official Journal; national implementations vary. We engineer controls and evidence — we are not a notified body, a certification body or a law firm, and we take no referral fees from any of them.
Nine controls carry four regimes
IEC 62443-4-1, the CRA, SOC 2 and ISO/IEC 27001 overlap far more than their table of contents suggests. Build these nine properly and most of the evidence pack assembles itself on every merge to main.
What each specification actually decides for your architecture
A working index, not a catalogue. For each reference: the concrete design decision it forces, and the artefact that proves you made it deliberately.
Reliability audit §2 — redacted, otherwise exactly as delivered
Ten working days against four services. Client name, hostnames, register maps and thirteen findings removed. Nothing rewritten for the website — including the finding we got wrong on first pass and corrected.
{{ f.detail }}
Technical buyers do not buy from a brochure. They buy from a working integration.
We do the engineering half of go-to-market: the reference architecture, the integration guide with real code, the benchmark methodology nobody can pick apart, and the demo environment that survives a sceptical staff engineer poking at it live.
What we produce
A launch, sequenced
{{ ph.body }}
Benchmark discipline
A performance claim is only worth the methodology behind it. Every benchmark we publish on your behalf states the hardware, the dataset, the version under test and the numbers we did not like.
Not our lane
Where we have shipped before — and what breaks there
Four sectors we know at protocol level. Outside these we will say so on the first call rather than learn your domain on your budget.
{{ ind.body }}
Four ways to buy. All of them fixed in writing before we start.
Ranges are real and current for US-based engagements. The number you get after a systems review is fixed for the quarter, and the review itself is unbilled.
{{ m.body }}
What is in the rate, and what is not
Describe the system and the constraint. We will name the model, the number and the start date on the call.
Five systems in production, with the numbers and the parts that went badly
Client names are withheld under NDA; every number below is measured, with its basis stated. Where a project overran or a decision was wrong, it says so.
{{ c.situation }}
{{ c.built }}
What actually shipped, week by week
A redacted extract from the portfolio log we keep internally. Client identifiers replaced with engagement references; nothing else edited.
How a sprint runs, and who is in the room when it does.
Two-week cadence, demo on the second Thursday, no status decks. Everything below is the default; we adapt to your ceremonies rather than importing ours.
The first eight weeks
A fortnight, in detail
Who decides what
Engineering standards we hold ourselves to
Cloud-style delivery, on iron we own, on American soil.
USA-DC1 is a small facility we own and staff in Ashburn, Loudoun County, Virginia — the densest fiber market in the country. You get the deployment ergonomics of a public cloud with a defensible answer to where the data physically is, who can touch it, and which jurisdiction governs it.
Where it sits, and why there
39.0438° N, 77.4874° WLoudoun County carries an outsized share of global internet traffic for one reason: fiber density. Two diverse conduit paths out of the building reach separate carrier hotels, which is what makes a 100 Gbps commitment meaningful rather than aspirational.
Three places your workload can live
{{ h.body }}
Data sovereignty, stated plainly
Most vendors answer this with a region name. Here is the whole chain, because that is what your customer's security team is actually asking about.
Customers and their auditors are welcome on site by appointment, escorted, with the access log available afterwards. Ashburn is thirty-four minutes from our Reston office and twelve from Dulles.
An American engineering company with its own iron and its own pager.
Uptime Systems AI, Inc. is a Delaware corporation with engineering offices in Reston, Virginia and Stamford, Connecticut, and a datacenter we own and operate in Ashburn. Twenty-six people, all US-based, all on the tools. No offshore delivery centre, no bench of subcontractors you meet after signing.
Who you actually work with
One of the eight below is on your first call, and stays on the engagement. We do not run a sales team.
Runs the company: which engagements we take, how pods are staffed, what we commit to in writing, and what happens when a commitment slips. Twenty-two years building and operating software for industrial and regulated environments before founding Uptime in 2023. He answers escalations himself, and he is the signature on the datacenter's operating policy.
{{ t.bio }}
Where we are
How the company is built
Deliberate constraints, all of which cost us revenue and none of which we intend to relax.
Engineers who carry the pager for what they ship
All roles are US-based and require the right to work in the United States without sponsorship; production access requires US-person status. Hybrid from Reston, Stamford or Ashburn, two days a week in the office, plus real travel to plant floors.
How we hire
The honest version
What we do with your code, your data and your access
Written for the security reviewer, not for the brochure. If your questionnaire asks something not covered here, we will answer it in writing rather than referring you to a portal.
{{ t.b }}
{{ t.ref }}Notes from the work, written by the people who did it
{{ a.dek }}
{{ art.title }}
{{ art.dek }}
{{ b.text }}
{{ b.text }}
The questions every prospect asks, answered without hedging
{{ f.a }}
Forty-five minutes with the engineer who would lead the work.
Not a discovery call with a salesperson. You leave with an architecture sketch, a sprint-one scope, and a straight answer on whether we are the right team for it.
Rory reads every enquiry himself and routes it to whoever should actually be on the call. If it is urgent, write to info@uptimesystemsai.com and say so in the subject line.
Legal notices, privacy and terms of use
Uptime Systems AI, Inc. is a corporation organised under the laws of the State of Delaware, United States, with engineering offices in Reston, Virginia and Stamford, Connecticut, and a datacenter in Ashburn, Virginia. Last updated 6 September 2026.
{{ l.h }}
{{ l.b }}
Agents that do one job, with typed tools and an audit trail.
We do not build a general assistant for your company. We build narrow agents with a defined task, a fixed set of typed tools, a hard turn limit and a human approval gate wherever money, safety or a regulator is involved. Every action the agent takes is logged with its inputs, its outputs and the model version that produced it, and every run is replayable.
Five agents we have shipped
{{ a.body }}
The runtime, drawn honestly
There is no magic in the middle. A planner selects from a registry of typed, idempotent tools; retrieval is scoped to documents the caller is allowed to see; a policy layer sits in front of every side effect; and everything lands in an append-only audit log that can replay the run.
Non-negotiable engineering rules
Failure modes we design against
SCROLL →Where the model runs
Model choice is a routing decision, not an identity. Each task is pinned to a model and a version, benchmarked against the alternatives on your own held-out set, and re-benchmarked before any upgrade.
Send us the task, two hundred real examples and the definition of a correct answer. Three weeks later you get an agent, an eval harness and an honest verdict on whether it should exist.