FDA AI Navigator
1 / 12

FDA AI Navigator

Explainable AI-powered regulatory intelligence for FDA-authorized medical devices.
Deterministic retrieval. Specialized agents. Every answer grounded in the FDA record.
Not a search box. Not a chatbot.

0
AI-enabled devices
0
companies
0
regulatory pathways
Jun 2026
data as of
Google AI Launchpad · live data ingested from FDA sources · Live App Demo · GitHub Repository · → to begin
The problem

The FDA record exists, but regulatory intelligence is fragmented across multiple systems.

510(k) databaseclearances, own search UI
PMA databaseapprovals, separate system
De Novo recordsno public API at all
Product classificationcodes, panels, regulation numbers
AI-Enabled Device Lista web page, updated on FDA's schedule
Has a device like mine already been authorized?
Which pathway did it use, and what does that mean for me?
Who else is operating in this space?
Answering these queries requires hours of manual cross-referencing across fragmented FDA sources, while generic AI answers lack verifiable grounding. Inspired by our prior radiology AI research, exploring how regulatory and evidence gaps bottleneck medical AI.
Why now

FDA AI authorizations grew 18× in a decade.

333 in 2025 alone, versus 18 in 2016. Radiology carries 76% of the catalog, but cardiovascular, neurology, and eight other specialties are climbing. AI-device authorizations are accelerating faster than regulatory teams can track manually.

*2026 is a partial year: data as of June 16. Computed live from our FDA data artifact, exactly as the product renders it.
Our position

Beyond Search. Toward Regulatory Intelligence.

Traditional search finds records. FDA AI Navigator explains, compares, and verifies.

LEVEL 1

Find

Search and filter 1,524 devices by specialty, pathway, year, and company: one unified catalog reconciled from every FDA source, instead of five separate databases.

FDA AI Navigator◐
Device Intelligence

Search and filter the catalog of 1,524 FDA-authorized AI medical devices.

All specialties ▾ All pathways ▾ Any year ▾ All applicants ▾ Software-only inferred
12 devices match
ContaCT
Viz.AI, Inc. · Radiology
De Novo DEN170073 ↗2018-02-13
Rapid NCCT Stroke
Ischemaview, Inc. · Radiology
510(k) K222884 ↗2023-03-02
LEVEL 2

Understand

Plain-language pathway explanations, market trend dashboards, competitive timelines: the regulatory meaning behind the records, not just the rows.

FDA AI Navigator◐
Market Intelligence

Approval patterns across 1,524 FDA-authorized AI medical devices.

Authorizations per year
19
20
21
22
23
24
25
26
By regulatory pathway
510(k)1,466
De Novo39
PMA19
LEVEL 3

Prove

Get answers with real FDA submission numbers, verified citations, confidence levels, and stated limitations, grounded in the official record before it reaches you.

FDA AI Navigator◐
↑
Analyzing your regulatory question...
✓ Identifying relevant FDA devices (12 matches)
✓ Evaluating regulatory pathways (risk class, then predicate)
✓ Validating evidence (3 verified)

ContaCT (Viz.AI) created this category via De Novo in 2018; later devices cleared via 510(k) as substantially equivalent.

DEN170073 ↗K222884 ↗K212261 ↗● High confidence
Who it serves

Three people, one shared bottleneck.

Different jobs, same hours lost to FDA archaeology.

⚙️

The Medical Device Startup

Needs regulatory precedents to identify predicate devices and map clearance pathways quickly.

🔬

The Researcher

Studies AI device trends. Needs a clean, traceable, and citable dataset.

🏥

The Clinical AI Evaluation Lead

Evaluates clinical AI deployments. Must verify vendor claims against the official FDA record.

⚙️ The Medical Device Startup

“What has already been approved?”

“We're building stroke-triage software. Who got there first, and how?”
  • Pathway precedent in seconds: filter the catalog to stroke AI and the history is right there.
  • The competitive field, as filed: every applicant, every decision date, straight from FDA records.
  • Shareable evidence: filters live in the URL, so a regulatory strategy discussion starts from the same view.
🔍 stroke · Device Intelligence
ContaCT (Viz.AI)
Viz.AI, Inc. · Radiology
De Novo
DEN170073 · 2018
Rapid NCCT Stroke
Ischemaview, Inc. · Radiology
510(k)
K222884 · 2023
Brainomix 360 Triage Stroke
Brainomix Limited · Radiology
510(k)
K232496 · 2023
Methinks NCCT Stroke
Methinks Software, S.L · Radiology
510(k)
K250685 · 2025
💡 The story a builder needs: one 2018 De Novo created this category: eleven 510(k) clearances have followed it. This precedent helps a startup understand the likely regulatory landscape and the devices a regulatory team may need to evaluate: 96% of all AI devices clear via 510(k).
🔬 The researcher

A reproducible dataset you can cite, not scrape.

“How is the market structured, and is entry still accelerating?”
  • Reconciled at ingestion: FDA's AI list cross-checked against openFDA bulk data, disagreements logged, never papered over.
  • Honest timelines: zero-count years stay on the axis; partial years are marked, not hidden.
  • Every row traceable: keyed by FDA submission number, linked to the source record.

By regulatory pathway

All 1,524 devices
1,524 devices
510(k) 1,466 · 96.2%
De Novo 39 · 2.6%
PMA 19 · 1.2%

New market entrants per year

First FDA authorization per company
2020
502021
612022
462023
852024
882025
1092026*
21
109 companies entered in 2025, the most ever. Entry is still accelerating. (Live from /api/trends, group_by=entrants.)
🏥 Clinical AI Evaluation Lead

Verify the vendor,
not the brochure.

“Three vendors claim FDA clearance for stroke triage. What does the FDA record actually say?”
  • Claims → submission numbers: every device in an answer cites its real FDA record.
  • Confidence, stated: high, medium, or low, with the reason why.
  • Limitations, admitted: what the data can't tell you is part of the answer.
“Which FDA-authorized AI tools triage large-vessel-occlusion stroke?”
Multiple AI devices are FDA-authorized for LVO stroke triage. ContaCT (Viz.AI) established the category via De Novo in 2018; subsequent devices including Rapid NCCT Stroke and Brainomix 360 cleared via 510(k), citing substantial equivalence within this landscape.
DEN170073K222884K232496K212261
● High confidence 4 of 4 citations verified against the FDA record
⚠️ And when the data can't support an answer, the system says “no verifiable answer”: a typed error, not an improvised paragraph. In healthcare, refusing to guess is a feature.
How it works

A multi-agent pipeline where every claim is checked.

Google ADK orchestrates specialized AI agents while deterministic services retrieve, enrich, and validate FDA evidence.

USER QUESTION “Which AI devices triage stroke, and how were they cleared?” ROOTAGENT · ADK ORCHESTRATION 🎯 Intent Agent interprets the question, extracts search parameters search parameters 🗄️ FDA Search Agent Deterministic SQL search over the local dataset: no scraping, no guessing TOOLS search_devices get_device · compare_devices get_trends · get_pathway_info Deterministic SQL query filters matching devices 🧬 Enrichment Agent joins classification, specialty, and AI category onto each device search_device_classification search_registration_listing enriched records ⚖️ Regulatory Analysis Agent explains pathways and precedent; it cannot invent retrievals draft + citations 🛡️ Evidence Validator deterministic Python: verifies every citation verified citations only ✓ DEN170073 ✓ K222884 ✓ K212261 Validated, cited answer real submission numbers · confidence level · stated limitations
The grounding rule: every citation is verified against the FDA record before the answer ships; a submission number that cannot be verified never reaches the user.
Architecture

System Architecture

Hybrid architecture: deterministic FDA services plus agentic reasoning built on one source of truth.

FDA AI-Enabled Device List the agency's official AI device list openFDA bulk downloads 510(k) · PMA · De Novo · classification planned production sync · run_sync.py next deployment phase a failed sync leaves the previous artifact serving read-only SQLite artifact 1,524 devices · Engine-Seam: scales to Postgres/Cloud SQL · GCS mirror (planned) SQLite ships inside the backend image FastAPI backend · Cloud Run one Python server, two execution paths DETERMINISTIC · NO LLM SQL over the artifact /api/meta · /api/devices · /api/trends instant, reproducible, zero hallucination surface AGENTIC · JUDGMENT REQUIRED ADK agents · Gemini Intent → Search → Enrichment → Regulatory → Evidence every citation gated by the evidence validator JSON JSON response with client-side progress simulation React frontend · Cloud Run Explorer · Trends · Pathways · Ask Navigator
React + Vite FastAPI SQLite artifact Google ADK Gemini Cloud Run Cloud Storage Cloud Scheduler
Interactive Walkthrough

The App in Action

Open Deployed Web App

STEP 1

Ask Navigator

The user submits a natural-language query directly from the dashboard landing page.

“Which FDA-cleared AI devices exist for stroke detection?”
STEP 2

Multi-Agent Pipeline

The query triggers the ADK reasoning graph. Users monitor agent execution steps in real-time.

✓ Identifying relevant FDA devices
✓ Evaluating regulatory pathways
✓ Validating evidence
⟳ Generating explainable report
STEP 3

Explainable Output

Every claim is synthesized, structured, and gated by the validator with links to official sources.

DEN170073 K222884
● High confidence
Future Vision

From FDA search to regulatory intelligence

Today

  • ✔FDA AI device catalog
  • ✔Ask Navigator
  • ✔Market Intelligence
  • ✔Explainable evidence

Next

  • →Semantic search
  • →FDA guidance RAG
  • →Cloud sync
  • →Real SSE agent execution

Future

  • →Regulatory Strategy Advisor
  • →Competitive Intelligence
  • →Similarity Search
FDA AI Navigator can evolve from a device-search product into an explainable regulatory intelligence platform.
scroll or ↑ ↓ · F fullscreen