The costly workflow
Health-oriented prototypes can create false confidence when polished outputs are presented without data provenance, failure cases, or explicit evaluation thresholds.
Research prototype
EKG Intelligence Lab
A research interface for testing signal-analysis ideas against public or synthetic EKG data, with failed evaluation gates preserved instead of hidden.
Researchers evaluating explainable signal-analysis workflowsActive research prototype · 2026Evidence reviewed July 28, 2026
Interface evidence
A real frame from the documented system
Each image is an optimized public-safe capture. Captions state the fixture, data, and verification boundary represented in the frame.
Patient-free evidence view showing aggregate public-dataset provenance and model gates, including an experimental gate that remains failed. Exploratory research only; not a medical device.
ProvenanceExisting Playwright capture of the EKG evidence workspace using public cohort manifests and aggregate evaluation results.
What Devin owned
Designed the research workflow, evaluation presentation, failure-state handling, and public-data safety boundaries.
System design
Architecture and data flow
The public description stays at a useful system level without exposing private records, credentials, or machine-specific implementation details.
Architecture
- Public or synthetic signal samples
- Experimental feature and scoring pipeline
- Evaluation gates and failure records
- Research-focused visualization interface
Data flow
- A public or synthetic signal enters the experiment.
- The research pipeline derives candidate features and outputs.
- Results are checked against explicit evaluation gates.
- The interface preserves both passing and failing evidence for review.
Constraints
- Only public or synthetic data belongs in the public demonstration.
- Failed evaluation gates must remain visible.
- The prototype is not a medical device and cannot provide diagnosis or treatment guidance.
System-flow diagram
How information moves through the verified slice
This is a public-safe architecture diagram, not a product screenshot. It shows responsibilities and handoffs without exposing private data or implementation paths.
Flow step 01
A public or synthetic signal enters the experiment.
Flow step 02
The research pipeline derives candidate features and outputs.
Flow step 03
Results are checked against explicit evaluation gates.
Flow step 04
The interface preserves both passing and failing evidence for review.
Evidence references
The artifacts behind the verification
Each reference names the implementation, test, documentation, or public demonstration that supports the claims on this page.
Pinned public-dataset manifest
Dataset versions, licenses, cohort metadata, and local integrity state are recorded separately from the public case-study package.
Signal and API test suites
Frontend and API tests cover signal-quality handling, supported inputs, analysis behavior, and service contracts.
Capability-gated model evaluation
Evaluation scripts keep model outputs unavailable or experimental when their recorded gates are not met.
Verified behavior
What the evidence supports
- The research package uses only public or synthetic data.
- Failed evaluation gates are documented as evidence, not removed from the narrative.
- Every public surface carries a clear non-diagnostic limitation.
The lab demonstrates a more responsible way to present early health-signal research: evidence and failure boundaries stay visible alongside the interface.
Boundaries
What this case study does not claim
- Not a medical device.
- Not validated for diagnosis, monitoring, or treatment.
- Experimental results do not establish clinical performance.
Proof-safety review
Public package checks
Confirmed for the material rendered on this case-study page and its linked public evidence.
- Confirmed — Uses only public or synthetic proof data
- Confirmed — Contains no secrets or credentials
- Confirmed — Contains no personal data
- Confirmed — Contains no machine-local absolute paths
- Confirmed — Contains no private records
- Confirmed — Exposes no writable production controls
- Confirmed — Contains no protected-system references
Implementation vocabulary
Signal processingResearch evaluationData visualizationPublic datasets