ApproachValidationGovernance Request a pilot
● engineering decision intelligence

A readiness score that can't be talked into anything.

Engineering Intelligence computes design readiness deterministically from your own evidence — then layers a strictly advisory, cited, local-only AI assistant on top that is architecturally incapable of changing the answer.

The governed decision chain — every readiness score, traced
Evidence
Coverage
Consistency
Validation
Confidence
Readiness

The problem

Every design review asks one question. Answering it is still manual.

"Is the evidence complete, consistent, and sufficient enough for a human to make an accountable decision?" Every regulated engineering team answers this by hand today — and every vendor is now racing to bolt an LLM onto that answer, which is precisely the wrong place to introduce uncertainty.

What "AI-assisted" usually means

A model summarizes your evidence, drafts a recommendation, maybe adjusts a risk score — and if it's wrong, there's no clean way to know which parts came from your data and which came from the model filling in gaps.

In a regulated design review, that's not a productivity feature. It's an audit finding waiting to happen.

What this platform does instead

The readiness score is a fixed formula over your own structured evidence — full stop. AI is a separate, clearly-labeled advisory layer that can summarize and cite, but is technically unable to move the number.

You can show an auditor exactly where every input came from, and exactly where the model's opinion started and stopped.

The approach

Two systems, deliberately kept apart

Most tools blur "the score" and "what the AI thinks" into one confident-sounding output. This platform draws a hard line between them, and the line is enforced in code, not just in the UI copy.

deterministic core

The readiness engine

Every score is computed from structured, human-entered evidence through a published, fixed pipeline. No model in the loop. The same inputs always produce the same result.

  • Evidence coverage against fixed regulatory categories
  • Structural consistency checks across findings and citations
  • Human-assigned confidence, not a model's confidence
  • A single documented formula for the final index
ECI = average(Coverage, Consistency, Validation, Confidence)
advisory layer

The AI review

One locally-hosted model, never a cloud call, producing schema-constrained output that must cite the evidence it's drawing from — or say plainly that it couldn't.

  • Every claim traced to a cited evidence ID
  • Provenance-hashed, so the exact request is reproducible
  • Provider failure falls back to a deterministic template — never a guess
  • Fixed "advisory only" label on every output, everywhere it appears
human_review_required: Literal[True] — enforced, not suggested

Why this is credible

Built like a regulated tool, not a demo

Four things that hold up under a skeptical technical read, not just a sales conversation.

Real regulatory grounding

IEC 62304, ISO 14971, IEC 60601, 21 CFR Part 803, and FDA design-control guidance are modeled as first-class reference data, not marketing copy.

Local-first by architecture

No external API dependency in the core workflow. The one AI integration point validates at startup that its endpoint is local-only — enforced, not promised.

Validated on real cases

Runs against actual public FDA recall cases alongside a synthetic composite, producing stable, explainable readiness scores across every one.

Process discipline as a product feature

Architecture decision records, a release-review checklist, and versioned release manifests — the kind of trail a compliance-minded buyer checks before the UI.

Validation

Tested against real regulatory history, not just synthetic data

Three case types, same deterministic pipeline, explainable results in every one.

Public FDA recall case

Tandem Mobi

A real, publicly documented device recall used to stress-test coverage and consistency scoring against actual regulatory evidence.

Public FDA recall case

Abiomed Impella CP

A second independent public case, confirming the scoring pipeline generalizes rather than being tuned to one dataset.

Synthetic composite

NovaPump

A controlled fixture case used as the platform's regression anchor — same formulas, fully reproducible, used in every automated test.

Governance

The rules aren't a policy page — they're architecture decisions

Two decisions, on the record, that any technical evaluator can go read.

ADR-002

AI is advisory

AI output must be structured, cited, limited, and visibly distinct from governed facts. It cannot create evidence, alter results, approve records, or declare compliance.

ADR-009

Local AI gateway

Provider-independent, schema-constrained, local-only by validated configuration. Provider failures reject execution — they never silently fall back to a guess.

Where it stands today

Stated plainly, not oversold

Pre-revenue · MVP

This is a single-user, local-only platform today — no authentication, no multi-tenant deployment yet, and it has been validated on public case data but not yet piloted inside a live design team. The durable audit-ledger and multi-specialist AI review layers already exist as designed interfaces, deliberately held back from the live product until the current single-specialist experiment is reviewed and accepted. That sequencing is a considered roadmap, not a missing feature. The raise stays deliberately lean: the AI runs locally, so there's no metered inference cost to fund — the ask covers design-team pilots, not a full go-to-market build-out.

Bring one real design review.

Run it against your own evidence and compare the output to what your review board would have concluded independently — or talk through where this fits as an investment or partnership.

This platform supports engineering judgment — it does not independently determine product safety, regulatory compliance, root cause, or corrective-action effectiveness, and does not itself constitute engineering approval.