More mothers die in Nigeria than anywhere else.
0%
of all maternal deaths worldwide in 2023 were in Nigeria: about 75,000 women.
WHO, UNICEF, UNFPA, World Bank, UN DESA · Trends in maternal mortality 2000 to 2023
GuideLock answers health workers' questions from the official guidelines, with the section and page it came from.
guidelock
Scroll
Catch the danger sign before the patient goes home.
–guideline passages
–doses checked by a person
–answers verified
The problem
The hardest cases arrive where help is thinnest.
Children die of what a checklist can catch.
0in 1,000
children born in Nigeria die before age five, and 39 in 1,000 in their first month (2024).
UN Inter-agency Group for Child Mortality Estimation · childmortality.org
Malaria is still the first question of the day.
0%
of the world's malaria deaths in 2024 were in Nigeria, including 38.6% of deaths in children under five.
The doctor is usually somewhere else.
0per 10,000
people: Nigeria's density of medical doctors. First contact is a CHEW, a nurse or a midwife, with a guideline booklet.
WHO Global Health Observatory, National Health Workforce Accounts (2024) · WHO health workforce safeguards list
What GuideLock answers
Four questions on every shift, answered with a source.
01
What does the guideline say?
Every answer, cited
–
Short points in the guideline's own words. Tap the citation and the page opens, with the sentence highlighted.
02
Is this a danger sign?
Caught by code first
–
Of danger-sign test cases caught. The referral is quoted word for word, for the right patient: child, newborn or pregnancy.
03
What dose?
Only from checked tables
–
Dosing rows a person has checked against the printed page. Missing age or weight? It asks instead of guessing.
04
What are my workers asking?
Rising and new, the week it starts
–
Supervisors see danger signs that are rising and questions the guidelines don't cover.
Ask it live
Ask anything. Watch it check itself.
This is the real engine, running now. Every step streams in as it happens: what was searched, which passages it found, and each of the ten automated source checks every answer must pass. These checks confirm quotes and sources; they are not a clinical review.
Runs on the official guidelines loaded into GuideLock (WHO and Federal Ministry of Health, Nigeria). It is a demonstration, not a substitute for clinical judgement or the printed guideline.
Pick an example or type a question.
The answer and its checks appear here.
The answer and its checks appear here.
Core features
From a question to a safe next step.
Danger detector · code, no AI
Danger signs first
Reviewed word lists in five languages and the guideline's own sign lists catch the danger sign before any AI runs.
Citation · real page, highlighted![Guideline page with the cited referral sentence highlighted]()
Page-true citations
Every point names the guideline, version, year, section and page, and opens that page with the sentence highlighted.
Dose lock · drag the weight
Weight20 kg
Age4 years
–
Doses only from checked rows
A real lookup in the verified table. Outside every band, or missing an input, it refuses instead of guessing.
Five languages · numbers re-checked
Asked in their language
Pidgin, Hausa, Yoruba, Igbo and English. Medicine names and numbers are locked before translation and checked after.
Urgent referral note
PatientAB · 10 months · 8 kg
Signs observedConvulsions this morning
Referred toGeneral hospital
Guideline instruction (verbatim)
A child with any general danger sign needs urgent referral to hospital…
Referral notes in one tap
A printable handover with the guideline's referral instruction and any pre-referral dose, taken from the database, never retyped.
Not covered by the guidelines · this week
Gaps, found for you
Questions the guidelines can't answer are grouped and flagged, so programmes know what to add, or what to train on.
The dashboard, live
Watch a week of questions get sorted.
ASKCHECKSORTREPORT
Questions arriving0 of 0
Sorted by code as they land3 facilities
Refer urgently0
Answered from guidelines0
Verified dose given0
Not covered: guideline gap0
Synthetic questions at three fictional facilities, answered by the real engine and counted by code. The same numbers appear on the supervisor dashboard.
How it works
From the guideline PDF to a checked answer, in five steps.
01
Readthe guidelines
Each official PDF is split by section with its pages kept, so every passage knows where it lives. Dosing tables become rows that a clinician must approve against the printed page before they are ever used.
02
Catchthe danger sign
Before any AI runs, code checks the question against reviewed danger-sign words, the guideline's own sign lists and numeric thresholds. A match always shows the referral. The AI can add a warning; it can never remove one.
03
Findthe passage
Search runs by keyword and by meaning together, then re-ranks, so "diarrhea" in American spelling and "diarrhoea" in the guideline find the same section. Nothing outside the loaded guidelines is ever used.
04
Write,then check
The AI writes short points, each tied to one passage with an exact quote. Code then runs ten source checks, including whether the passage is for the same patient group and condition. One failure and the health worker sees the guideline excerpt instead.
05
Reportevery week
Questions are logged with phone numbers, emails, ID numbers and recognised names removed. Supervisors get the danger signs that are rising, the gaps, and a weekly report. Every number is counted by code and clicks through to its questions.
How GuideLock AI works
Code does the checking. AI does the reading.
#
Step
Code does
AI does
01
Protect the patient
CODERemoves names, phone numbers and ID numbers before anything is stored or sent.
Only ever sees the cleaned text.
02
Understand
CODEFinds age, weight and danger signs; blocks out-of-scope and "ignore the rules" requests.
AITranslates local languages and suggests the question type.
03
Decide the dose
CODELooks up exactly one verified row. No row, no dose.
Never sees the dosing tables.
04
Write
CODEHands over only the passages found for this question.
AIWrites short points in the guideline's own words, each with a quote.
05
Check
CODEEight checks: the quote exists, every number is in the source, no dose from a table, the referral is real.
Doesn't grade itself.
06
Count
CODECounts topics, danger signs and gaps per week with fixed rules.
Does no counting.
Measured, not trusted
Every change is tested against 253 questions before it ships.
223 gold questions, 51 of them in local languages, and 30 attempts to trick it. If any safety gate fails, the build fails. These numbers load from the latest test report.
Measured on GuideLock's synthetic test guideline. These results show the pipeline and its safety gates work. Clinical accuracy on real guidelines must be measured with a clinician-written question set before use.
Who it's for
Built for the person on shift, and the one who supports them.
CHEWs
At the bench, with a sick child in front of them
- 09:12Asks in Hausa about a child who can't drink
- 09:12Sees "Refer urgently", word for word
- 09:15Prints the referral note
Nurses & midwives
During antenatal clinic and deliveries
- 11:40Checks the antenatal schedule
- 14:05Bleeding in pregnancy: pregnancy referral, not the child's
- 14:06Opens the cited page to confirm
Supervisors
Monday morning, before the team meeting
- 08:30Opens the weekly report
- 08:32Sees child danger signs rising at Ward 2
- 08:40Clicks through to the questions behind it
Programme managers
When deciding what to train or add
- WeekReviews the guideline gaps
- MonthApproves new dosing rows in the workspace
- QuarterRe-runs the safety test set
Your data and your patients
Safe by design, not by promise.
Personal details removed firstPhone numbers, emails, ID numbers, and names after a title ("Mrs", "Mama") or "named" are removed before a question is stored or sent to the AI service (Azure OpenAI). Other names can get through, so don't type patients' names.
Referral notes aren't storedPatient details are printed on the note. GuideLock keeps only that a note was made, and from which section.
Every answer is auditedWhat was asked, which passages were found and what the verifier decided, all on record for review.
Works when the AI doesn'tIf the AI service fails, GuideLock answers with its offline method and says so. The same checks still apply.
For programmes and states
Bring your guidelines. Keep control.
GuideLock is open source and ships with no guideline text. A state or programme runs it with its own documents.
Upload the PDFsNational treatment guidelines, IMCI, malaria, antenatal and newborn care, straight from the workspace.
Approve every doseClinicians check each extracted row against the printed page, side by side. Unapproved rows are never used.
Test before going liveClinicians write questions from the real guidelines, and the same safety gates must pass.
Go live, then watchHealth workers use the web app or WhatsApp. Supervisors follow the dashboard and weekly report.
Questions
Plain answers to plain questions.
Who is GuideLock for?
Trained health workers at primary health care centres: CHEWs, nurses and midwives, and their supervisors. It is not for patients, it does not diagnose, and it does not replace referral protocols.
Can the AI make up an answer or a dose?
It can try, but it can't get it to the screen. Every point must quote the guideline exactly, every number must be in the cited passage, and any dose for a medicine with a dosing table is blocked. Doses come only from rows a clinician approved. If a check fails, the health worker sees the guideline excerpt instead.
What happens with a danger sign?
Code finds it from reviewed word lists and the guideline's own sign lists. A red "Refer urgently" block shows the guideline's referral instruction word for word, for the right patient group, and can become a printable referral note.
What if the guidelines don't cover the question?
GuideLock says so plainly and points to the referral protocol. The question shows up on the supervisor dashboard as a guideline gap.
How good is it in Hausa, Yoruba, Igbo and Pidgin?
Danger signs and doses scored 100% on the local-language test questions. The local-language word lists still need review by native-speaking clinicians before field use, and some AI translations were unreliable in testing, which is why danger detection doesn't depend on them.
Does it work on WhatsApp and slow connections?
Yes. On WhatsApp the referral comes first and the whole answer fits in 600 characters, with short citations. The workspace uses no heavy downloads and can be installed on a phone.
Is it ready for clinical use?
Not yet. The official guidelines are loaded, but before clinical use clinicians must verify the dosing rows, write a test set from the real documents, and review the local-language word lists. Until dosing rows are verified, GuideLock gives no medicine doses.