Refuses unsafe drafts
Discharge stays blocked while pending results lack owners. Allergies and medication reconciliation must be explicit at admission. Enforced in code, not suggested.
Nothing listens. Nothing leaves. You stay the author.
MedSeek turns the notes you already type into checked SBAR handovers, SOAP notes, admission H&Ps, and discharge summaries - blocking unsafe drafts until gaps have owners, scoring patient instructions against a 6th-8th grade target, and citing every lookup with source and retrieval date.
Requires DeepSeek Harness (free, open source) and Node 22 or later. Restart the web GUI after install.
Runs locally
Apache-2.0
No telemetry
Hash-chained audit trail
De-identifies first
HIPAA Safe Harbor category scrubbing with published corpus numbers and an explicit residual-risk statement. It assists de-identification; it does not certify it.
Auditable by construction
Every tool execution writes a content-free hash-chained receipt you can verify yourself. No note text, no queries - digests only.
Three shifts, one minute to review
Inpatient staff nurse
12-hour shift, 4 to 6 patients. Handover at both ends — paste flowsheet rows, get an SBAR or I-PASS scaffold, completeness check before you sign. No re-typing from scattered charts.
Weekly: SBAR at handover · MedlinePlus handout before discharge teaching
Hospitalist / resident
Census 12 to 18, daily admits and discharges. Discharge summary refuses to pass while pending results have no owner — fixing the 65% hole Kripalani described. I-PASS for night cross-cover.
Weekly: discharge drafts daily · label lookups on rounds · PubMed for teaching
Outpatient GP
20 plus visits at 15 to 30 minutes, portal inbox that never empties. Patient instructions back to 6th to 8th grade plus an NLM handout in one turn. No portal paste into chatbots required.
Weekly: instruction rewrites · patient education lookups · PubMed checks
All three share one trait from the PRD: they will use a tool weekly only if it starts from text they already have and returns something they can review in under a minute.
Evaluation
Measured on a published corpus, not vibes
95.9%recall*
100%precision
100%detection
0false alarms
*synthetic corpus - regenerate: pnpm eval · Read the eval report →
Measured on a published corpus, not vibes
On our fully synthetic annotated evaluation corpus: identifier recall 95.9%, precision
100%; completeness detection 100% with zero false alarms. Regex detection fails silently
in the dangerous direction - so the tool labels uncertainty instead of claiming
certification. Regenerate the numbers yourself: pnpm eval.
Read the eval report →
Paste, scaffold, validate — gaps in plain text
Every drafting tool does the same two-call rhythm. First call returns the checklist and guidance. You draft each field from text already in the conversation. Second call validates: missing required fields, unresolved pending markers, error-prone abbreviations, and an assembled draft with a review banner and a next-step line.
Lookup tools do the inverse: they return source, identifier, link, and retrieval date so the citation survives copy-paste. Empty means no results, never a filled gap.
$ sbar_handover_draft({format:"sbar"}) → scaffold
SBAR structure (IHI SBAR):
Situation (required): one-sentence reason for handover.
Background (required): relevant history and context.
Assessment (required): problems addressed and risk descriptively.
Recommendation (required): what is needed and who owns it.
$ sbar_handover_draft({filled_fields:{...}}) → validated
DRAFT — review required
S: Rapid response, 68M post-PCI day 2, hypotensive
B: Afib on apixaban, last troponin 1.2
A: Suspected bleed — Hgb 7.1 down from 9.4
R: Hold next dose pending attending, recheck Hgb 0600
Gaps (0): Structural check passed.
Next: Present the draft to the clinician for review.
Drafts, never directives
MedSeek restructures clinician-supplied text and retrieves public reference material. A named clinician reviews everything before it is used. It does not diagnose, prescribe, compute doses, rank differentials, score drug interactions, or assign billing code levels. It is not a medical device. See the enforcement model →
BLOCKED
pending result - no owner
PASSED
owner assigned - summary completes
owner assigned - summary completes
Local by design
The drafting, checklist, and de-identification tools make zero network calls. Lookup tools send terms and codes only - never identifiers - to exactly nine allowlisted hosts, and an egress guard denies calls whose arguments look like they carry patient identifiers. No telemetry anywhere in the bundle.
- eutils.ncbi.nlm.nih.gov
- pmc.ncbi.nlm.nih.gov
- www.ebi.ac.uk
- api.fda.gov
- rxnav.nlm.nih.gov
- dailymed.nlm.nih.gov
- connect.medlineplus.gov
- clinicaltables.nlm.nih.gov
- clinicaltrials.gov
New in 0.1.3
Feynman parity for the clinical reference surface
Eight additional narrow, citable lookup modes - PMC ID conversion, related articles, Europe PMC search and open-access full-text sections, FAERS adverse events, FDA recalls, Drugs@FDA applications and counts plus pharmacologic classes and generic equivalents, and ClinicalTrials details, eligibility, investigators, and endpoints. Nine allowlisted hosts, twelve lookup tools, seven drafting tools. Read the gap audit →