Dei Corpus
Three signal classes, one computed output.
Dei Corpus builds a multimodal fusion engine: software that explores correlation across signal streams from separate third-party devices. It runs on synthetic data only.
- Development start
- May 2026
- Stage
- TRL 4, technology validated in lab
- Data
- Synthetic only
- Availability
- Not available for supply
Problem statement
Good signals, separate readings.
Cardiovascular, metabolic and dermatological signals are captured well today. Consumer and clinical devices record cardiac time series continuously. Laboratory panels quantify metabolic chemistry at discrete points, and dermatological imaging is routine. Each stream is then read inside the system that captured it, against that system's own reference range.
Correlation across these domains does exist. Where it exists, it sits either in consumer wellness products outside the regulated pathway, or inside vertically integrated clinics that build their own sensors and operate their own premises. We are not aware of a regulated multimodal system that performs this correlation on top of already-certified third-party devices.
That is what we know, not what is true of the world. If such a system exists, write to us and we will correct this page.
- Cardiovascular
- Continuous time series, high sampling rate, captured by devices a person already wears.
- Metabolic
- Discrete laboratory chemistry at irregular intervals.
- Dermatological
- Images, under capture conditions that vary by operator, device and lighting, on a schedule that is not linked to either of the other two streams.
The signals are already good. What is missing is a common frame in which to read them together.
Architecture
A fusion layer over certified inputs.
We are building a multimodal fusion layer designed to take input from already-certified third-party sensors. We do not manufacture hardware and we do not intend to.
The design assumes every input signal reaches the engine from a device that already carries its own CE marking or FDA clearance for its own intended purpose. That certification belongs to those devices and to their manufacturers. It does not extend to this software, which carries no CE marking of its own.
The engine takes inputs with different sampling rates and different units, with different missingness. It aligns them on a common time base and computes correlations across classes. Each class is encoded separately, and the output object carries per-class provenance. Missing data is carried through as an explicit state and is never silently filled.
Fusion
Cross-class correlation. Our scope.
Alignment and encoding
Common time base, per-class encoders.
Certified inputs
Third-party devices, third-party conformity.
Three in, one out.
- Capital
- No sensor manufacturing, no clinic real estate, no acquisition hardware to certify.
- Regulatory
- Any conformity assessment we pursue would be scoped to the fusion layer, not to the acquisition hardware. The hardware is certified by its own manufacturers under their own intended purposes, and that certification does not carry over to this software.
On other companies
We do not name competitors on this site. Under Directive 2006/114/EC a comparison must objectively compare features that are material, relevant, verifiable and representative. A pre-market company with no clinical results cannot verify a comparison against certified devices, so any such comparison here would be unlawful, and unearned.
What we can state without naming anyone: preventive whole-body scan services reach multimodal coverage by controlling the whole acquisition stack, whether through purpose-built sensors or through proprietary acquisition protocols running on commercial scanners, inside a fixed clinical setting. Clinical AI platforms are consolidating within hospital imaging and acute-care workflows. The input set we are researching is different, being streams from third-party devices used outside the radiology department.
Development status
Where the work actually is.
Development started in May 2026. The engine is at TRL 4 on the Horizon Europe scale, technology validated in lab. Validation has been carried out in a controlled laboratory setting on synthetically generated data. No real-world data, no clinical data and no personal data have been used, and no validation in a relevant or operational environment has been performed.
Results are measured against labels we generated ourselves, so they describe how the software behaves on simulated data. They are not evidence of clinical performance.
- Development start
- May 2026
- Stage
- TRL 4, technology validated in lab
- Data
- Synthetic only
- Clinical investigation
- None conducted, none authorised
- Conformity assessment
- Not started
- CE marking
- None
- Availability
- Not available for supply. No download, no API, no pilot.
- Team
- One founder
- Location
- Portugal
What does not exist yet
No clinical data and no clinical results. No clinical performance figures, and none published anywhere. No CE marking, and no notified body engaged. No signed hospital, research or commercial agreement. No customers, no revenue, no deployments. No employees beyond the founder.
Near-term objectives
Q4 2026 · Freeze the v0 input contract
One schema per signal class. One versioned output object. Explicit units, sampling rates and missing-data states, written down before more code.
Q4 2026 · Extend the synthetic generator
Once the contract is frozen, specify cross-class covariance explicitly, so the fusion layer can be tested against a correlation structure we injected ourselves. Publish the generation parameters so a third party can reproduce a run.
Q1 2027 · Open the regulatory file
Draft the intended-purpose statement. Write the classification rationale under Rule 11 of Annex VIII, Regulation (EU) 2017/745. Run a gap analysis against IEC 62304 and ISO 14971. Decide the conformity route before any data conversation starts.
Open · Seek one research collaboration
Retrospective, de-identified data, under an ethics-approved protocol and a data processing agreement. Nothing is signed, and this page will say so until something is.
Targets are internal planning dates, not commitments.
Roadmap
Four phases. Phase 1 is current.
The order carries information. Each phase depends on data access, regulatory scope and capital that the previous phase is meant to establish. Only Phase 1 has work in it today, and phases 2 to 4 carry no dates, because dates there would be fiction. Every phase is gated on conformity assessment under Regulation (EU) 2017/745. Nothing described here reaches clinical use before that assessment is complete.
Phase 1Current
Cross-class fusion on synthetic records.
Fusion across cardiovascular, metabolic and dermatological signal classes, on synthetic records, in a laboratory setting. All current work sits here.
Phase 2Not started
Medical imaging already acquired in routine care.
Extend the engine to imaging acquired on third-party certified equipment. No work has begun.
Phase 3Not started
Oncological liquid biopsy.
Add laboratory assay outputs as a further signal class. No work has begun.
Phase 4Not started
Neuro-oncology.
One line until Phase 3 produces something.
Contact
Write to the founder directly.
- Founder
- Lázaro Ramos
- me@lazaroramos.com
- Location
- Portugal
Do not send personal health data or patient data by email. This site processes no health data and we do not want any.
This link opens your own email client. Nothing is sent through this site.