Signals from the other side.
Field notes, methods and announcements from a studio that keeps receipts.
Why “impossible” is usually a data problem
Almost every brief we're told can't be done fails for one of three fixable reasons. Here they are — with the fixes.
Read the essay → 6 minThe velocity playbook: six weeks to regulator-ready
The exact weekly cadence behind our fastest regulated delivery.
Read the method → 4 minViceron joins the NVIDIA Inception Program
What accelerated computing changes about which problems are worth attempting.
Read the news → 2 minField notes: finding one patient in eight million
A diary of the eleven weeks everyone said would be wasted.
Read the notes → 8 minViceron and HPE expand on-premise AI partnership
Sovereign AI for organizations whose data can never leave the building.
Read the release → 2 minEnterprise AI governance: audit trails or it didn't happen
What boards and regulators actually accept — and how to build it in from commit one.
Read the essay → 5 minForecasting 40 hospital wards without moving a record
Federated learning in a public healthcare region — capacity planning with zero data pooling.
Read the notes → 5 minViceron on stage: "Engineering the impossible" — Copenhagen
A talk on why enterprise "can't" is usually a tooling budget, with live field examples.
Event details → 1 minOne email. Only when something ships.
Writing about us? We answer fast.
Interviews, background briefings and comment on applied AI in regulated industries. Brand assets and boilerplate on request.
Why “impossible” is usually a data problem
In ten years of collecting refused briefs, we have heard “it can't be done” for exactly three reasons. None of them is physics.
Reason one: the data exists but nobody can read it. Eight million records in forty formats is not a wall — it is a parsing budget. The answer was always in row 7,344,102; what was missing was the machinery to get there before the deadline.
Reason two: the signal is real but rare. Rare things defeat dashboards and averages. They do not defeat weak supervision, domain heuristics and models built specifically to learn from near-misses.
Reason three: the timeline was priced for a committee. Most eighteen-month roadmaps contain about six weeks of engineering. The rest is coordination. Remove the coordination, keep the engineering, and “impossible by Q3” becomes “live by Friday.”
The pattern under all three: impossible is rarely a property of the problem. It is a property of the tooling, and tooling can be built. That is the entire business model.
Next — The velocity playbook →The velocity playbook: six weeks to regulator-ready
Week one ends with software in production. Everything else follows from that single, non-negotiable rule.
Weeks 1–2: ship the skeleton. A thin end-to-end system — real data in, real decision out, audit trail on. Ugly is fine. Live is mandatory.
Weeks 3–4: evidence loops. Every week closes with a measurable claim: precision on review, hours saved, findings survived. Claims that fail die young, which is exactly when dying is cheap.
Weeks 5–6: harden and hand over. Compliance evidence has been accumulating since day one as a by-product of the pipeline — so the final fortnight is polish, not paperwork archaeology.
The trick is not working faster. It is refusing to let anything exist that cannot be demonstrated. Demos don't slip; documents do.
Next — Joining NVIDIA Inception →Viceron joins the NVIDIA Inception Program
Viceron is now a member of NVIDIA Inception, the program for companies advancing AI and data science.
In practice it means our experiments run on accelerated computing that turns week-long training runs into overnight ones — which changes not just how fast we work, but which problems are worth attempting at all.
A rare-cohort model that needs fifty iterations to converge is a non-starter at one iteration per week. At five per night, it is a Tuesday. Compute budgets quietly define the boundary of the possible; ours just moved.
Together with our HPE infrastructure partnership and Danoffice IT on global hardware, the stack under the impossible is now very much our own.
Next — Field notes: one in eight million →Field notes: finding one patient in eight million
Week zero: the client's own registry team has formally concluded the patients are not findable. We take the brief anyway. This is the diary.
Week 1. The diagnosis codes are wrong more often than right — as warned. But prescriptions, lab panels and referral trajectories tell a story codes don't. The magnet will be built from side-effects of the truth.
Weeks 2–5. Weak supervision: every rule a clinician can state becomes a noisy label. The model learns from thousands of almost-cases. Precision on the first review round: embarrassing. On the third: interesting. On the fifth: quiet in the room.
Weeks 6–10. Every candidate now ships with an evidence trail a clinician can argue with. The arguing is the point — each round of it retrains the magnet.
Week 11. Two hundred and fourteen candidates, ninety-six percent confirmed on review. Somewhere in row 7,344,102, a patient who was officially unfindable. The board meeting goes very quiet.
First — Why impossible is a data problem →Viceron and HPE expand on-premise AI partnership
Copenhagen — Viceron today announced an expanded infrastructure partnership with Hewlett Packard Enterprise, bringing sovereign, on-premise AI delivery to clients whose data cannot leave their perimeter.
Why it matters. Hospitals, public authorities and pharmaceutical companies increasingly need frontier AI capability without cloud data transfer. Under the partnership, Viceron systems deploy on HPE infrastructure inside the client's own data center — same models, same velocity, zero data movement.
What clients get. Reference architectures for regulated AI workloads, GPU-dense configurations sized with Danoffice IT, and Viceron's six-week delivery method running end-to-end on hardware the client owns and auditors can inspect.
The first joint deployments are live in Danish public healthcare. Enquiries: contact@viceron.com.
Next — Enterprise AI governance →Enterprise AI governance: audit trails or it didn't happen
Every enterprise AI program eventually meets a person with the power to stop it: an auditor, a regulator, a board risk committee. What survives that meeting is never the demo — it's the paper trail.
Governance is an engineering property. Model cards, decision logs, human-oversight checkpoints and reproducible training runs are features. If they're built into the pipeline from the first commit, compliance evidence accumulates for free with every deploy. If they're bolted on later, they cost more than the model did.
The EU AI Act changes the default. For high-risk systems, traceability and oversight stop being best practice and become market access. Organizations that treat this as product architecture, not paperwork, will ship while competitors write memos.
Our rule of thumb. A system you can't explain is a system you can't defend — and a system you can't defend is a system you can't keep. We build the explanation machinery first. It's also, not coincidentally, why our audits pass.
Next — Forecasting 40 wards →Forecasting 40 hospital wards without moving a record
The brief from a public healthcare region: forecast capacity across forty wards in eleven hospitals — where patient records legally cannot be pooled into one database.
The federated answer. The model travels; the data stays. Each hospital trains locally inside its own perimeter, and only model updates — never records — are aggregated centrally. Privacy review signed off because there was nothing to sign away.
What surprised us. Local training exposed local truths: two "identical" wards had opposite weekend patterns that a pooled model would have averaged into fiction. Federation wasn't just the compliant option — it was the more accurate one.
The result. Region-wide capacity forecasts on infrastructure the hospitals already trust, with every record exactly where the law expects it to be.
Next — Copenhagen event →“Engineering the impossible” — live on stage
A 30-minute talk on why enterprise “can't” is usually a tooling budget in disguise — with live walkthroughs of the eight-million-record needle and the six-week regulator-ready build.
For. Executives and technical leaders in pharma, logistics, healthcare and the public sector who keep hearing "impossible" from vendors.
Details. Date and venue announced to dispatch subscribers first. Seats are limited; enterprise teams can request a private briefing instead.
Reserve interest: contact@viceron.com with subject “Copenhagen talk”.
First — Why impossible is a data problem →