One agent per regulatory job.
Live today: 510(k) predicate research. Every Keenr agent is built the same way — narrow retrieval over the right corpus, evidence-first reasoning, outputs you can audit. No general-purpose chat.
Software is being rewritten as services-as-software.
In March 2026, Coatue and Sequoia independently published the same thesis: the unit of value is shifting from per-seat software to per-output work. Harvey brought it to legal research, TaxGPT to tax advisory, Anterior to healthcare revenue cycle — rule-governed work, deep domain corpus, evidence you can audit.
Medtech regulatory research sits in the same quadrant. Keenr is the medtech instance.
Three things general AI can't do for regulatory work.
Specialized beats general.
A generic chat model can't reason about substantial equivalence the way an agent purpose-built on the full 510(k) corpus can. Every Keenr agent is built for one domain, with the taxonomy, policy context, and evidence conventions of that domain already in hand.
Evidence beats assertions.
Every claim in a Keenr output links to its source: an actual 510(k) summary, an FDA guidance section, a MAUDE record. You audit the reasoning, not just the answer. Nothing is ever "because the AI said so."
Software speed beats manual cadence.
What used to be a multi-week research project compresses into a session you can rerun whenever the device or the landscape changes. The knowledge lives in the tool, not in a one-off deliverable.
One agent live. One in private beta. More coming.
Each agent compresses a discrete piece of regulatory or commercial research, work that today takes weeks of manual analysis, into a single session. AI/ML depth (Predetermined Change Control Plans, foundation model disclosure patterns, algorithm change precedents) runs across the full platform as a signature capability, not as a separate vertical.
Keenr Predicate Finder
510(k) regulatory research in a single session. Reads every 510(k) summary, reasons over substantial equivalence, scores against FDA's 2023 Best Practices guidance, drafts the 510(k) Summary narrative.
$499 / report · Try →
Cybersecurity Package Generator
FDA 2023 cybersecurity guidance package, SBOM templates, threat modeling.
Pre-Submission Strategist
Q-Sub meeting prep, pathway analysis, precedent search across FDA correspondence.
Letter-to-File Engine
Post-clearance change decisions, LTF-vs-Special-510(k) analysis, audit-ready memos.
Reimbursement Pathway Analyzer
From clearance to coverage: CPT/HCPCS and ICD-10 candidates, Medicare pricing across all three fee schedules (PFS, OPPS, DMEPOS), payer coverage outlook, novel-technology pathways, and an A–F reimbursement-readiness grade. Built. Running with design partners.
- · Competitor Clearance Monitor
- · Continuous post-clearance MAUDE monitoring
- · US→EU Crosswalk
- · Launch Readiness Playbook
- PCCP library: 80+ cleared Predetermined Change Control Plans extracted from 510(k) summaries, queried by semantic similarity on every AI/ML report.
- Algorithm change precedents: modification frequency analysis by category (training data, model retraining, performance specs, deployment), sourced from real cleared filings.
- Foundation model disclosure tracker: cleared AI/ML devices indexed by foundation model family, refreshed weekly via auto-ingest cron.
- AI/ML guidance RAG: 1,200+ chunks of FDA AI/ML guidance documents, semantically retrieved per device. Daily corpus-health smoke test guards retrieval quality.
- Citation freshness: every cited guidance, precedent, and disclosure carries an "Indexed YYYY-MM-DD" timestamp on the report, so reviewers can see when each was last indexed.
Applied across every agent above. Not a separate vertical.
The regulatory surface area the agent actually reads.
-
91K+ indexed 510(k) summaries
Every publicly available 510(k) summary is extracted, embedded, and searchable by semantic similarity, not just by product code — and the corpus refreshes every two weeks. Ranking holds even for the one-in-three product codes with zero prior clearances, territory keyword search cannot see at all.
-
Full openFDA clearance history
Queried live: structured 510(k) records, MAUDE adverse events, and recall data. Covers 180K+ clearances historically, fresh with each query.
-
FDA guidance documents
September 2023 Predicate Best Practices guidance, January 2025 AI-Enabled Device Software Functions guidance, and the supporting policy documents that shape substantial-equivalence reasoning. All chunked, embedded, and semantically retrievable. The agent cites what it's anchored to, by section.
-
PCCP & Foundation Model libraries
Two specialized libraries built on top of the 510(k) corpus: 80+ cleared Predetermined Change Control Plans (extracted, embedded, semantically searchable) and a foundation model disclosure tracker that watches how AI/ML devices describe their models. Both auto-refresh weekly.
-
De Novo grant library
Hundreds of De Novo grants embedded and retrievable. Novel-mechanism devices get matched De Novo precedents in the report, labeled by match strength — never overstated.
-
Citation freshness, on every page
Every guidance citation, PCCP precedent, and foundation model disclosure renders with an "Indexed YYYY-MM-DD" timestamp on the PDF. A daily smoke test asserts retrieval quality stays above threshold. Reviewers see when the corpus was last validated, not just what it says.
Summary corpus includes the boleary.com dataset (CC BY 4.0).
Retrieval, then a multi-step agent loop.
The corpus
- 91K+ 510(k) summaries
- openFDA · queried live
- FDA guidance RAG
- PCCP + FM libraries
- De Novo grants
The agent loop
- 01 Classify device + software posture
- 02 Score vs FDA Best Practices
- 03 Gate: clinical function + relevance
- 04 8-dimension SE comparison
Every step outputs evidence
The deliverable
35+ page report
Citation-linked · auditable · re-runnable
It finds predicates keyword search misses — matching by meaning first, product code second. Every candidate comes with its clearance date, applicant, product code, and full summary text.
Then the agent loop: classify the submitted device (including its software posture — SaMD vs SiMD, AI subtype, locked vs adaptive), score candidates against FDA's four predicate best practices, cross-check safety (MAUDE + recalls), and run an 8-dimension substantial-equivalence comparison. Candidates that don't share the device's clinical function are dropped, and nothing below the relevance floor is shown — so "no adequate predicate" is a finding you can trust, not a padded table.
The output is a drafted regulatory analysis: ranked candidates, primary recommendation, SE comparison table, pathway decision, testing implications, and a 510(k) Summary narrative you can adapt. If the classification is PMA-only, the report says so — no forced 510(k) framing — and maps the closest PMA and De Novo approvals in your code family instead. Everything is citation-linked back to the underlying 510(k), guidance section, or openFDA record. Nothing is trust-me.
Not a generic AI wrapper.
Keenr is led by Vinod (Vinny) Kaimal and built with practicing RA leaders, consultants, and clinical researchers who validate every agent against real workflows.
The posture: an operator's tool, priced so an RA lead can run one without opening a purchase order. Partner plans for consultancies and RA teams — editable Word delivery included — are available today.
Read the full founder storyReady to run the first agent?
$499 per report. No account required for the first search.