Too few health workers, too little signal, staff who are not there and forms that eat the visit. Open data for about 195 countries, with Indonesia and India in focus.
Noor lives two hours from the nearest clinic. When she gets there, she joins one queue for everyone. Here is why that queue moves so slowly, why Indonesia is ready to fix it, and why Indonesia and India are worth it, in seven numbers.
Presenting? Use ↓ ↑ to move between steps.
1 · Too few hands
DataWHO Global Health Observatory. Doctors plus nurses and midwives, latest value since 2015.
2 · Rural means offline
The more rural a country, the less of it is online. A tool that needs the internet fails exactly where care is scarcest.
DataWorld Bank WDI: rural population and internet users, latest since 2015.
3 · The roster is not the room
Queues punish every missing person. With 10 patients an hour and 15-minute visits, losing one of four staff makes the wait seven times longer. Lose two and the queue never clears.
LiteratureChaudhury et al., provider absence surveys. IllustrativeTextbook queue model; real mornings are worse.
4 · Paperwork eats the visit
Clinics keep up to 48 registers. Add absence on top, and of 100 rostered clinical hours only about 42 reach a patient.
LiteratureSiyam et al. 2021, five countries. IllustrativeCombined with absence averages.
5 · The rails exist, the staff time does not
DataWorld Bank WDI electricity, internet, mobile; WHO GHO workforce. LiteratureIndonesia only: Permenkes 24/2022, Kemenkes SATUSEHAT and SITB, Sahabat-AI. National averages, not every clinic.
6 · Two countries, most of the need
DataWorld Bank WDI population, rural population, TB incidence; WHO GHO workforce. Missing health workers = (44.5 − density) × population / 10,000. A headcount view: it shows where one tool reaches the most people, not where need per person is highest.
7 · Nobody can see capacity
Median record completeness is 12.5%. Without hours or capacity, nobody can send a patient to a clinic that is open and staffed.
Datahealthsites.io / OpenStreetMap (ODbL). The map shows the 7,448 facilities mapped as points; 10,414 more are building outlines with no coordinates in this export.
So the problem is
Because of Vitamin Bob, a rural patient with any phone will learn in one free call whether to call 108, go to the nearest open clinic now, or come at a booked time, instead of travelling to a clinic that may be closed and waiting in one first-come queue; we know because:
Built and tested first in India, in Hindi and Gujarati, where the scale is. Indonesia is the readiness case for the next language pack: 73% online, 99.9% electricity, and electronic records already mandatory in every puskesmas. Together the two hold 74% of the people in countries that are ready but short-staffed.
Problem statement in the hackathon template. See how Vitamin Bob works →
The product
No app, no internet, no cost to the patient. A small AI model on a district laptop listens in Hindi or Gujarati. Fixed rules decide the urgency, and a clinician confirms anything that is urgent or uncertain. Here is one real test call, step by step.
A mother's son has had a fever for two days and is breathing fast. She has a basic phone and no data. She rings the district number and hangs up before it connects, so the call costs her nothing.
For the districtThe missed call is logged in seconds and queued for a callback. Clinicians use the same trick to say "I'm on duty".
Stuart, the phone module, calls her back from the district line. A recorded menu asks her to press 1 for Hindi or 2 for Gujarati, then who is sick.
Safety by designEvery sentence she hears is a clip recorded in advance and written by the team. The AI never speaks to the patient.
"मेरे बेटे को दो दिन से बुखार है और खाँसी भी है, साँस बहुत तेज़ चल रही है।""My son has had a fever for two days, and a cough too. He is breathing very fast."
Gemma 4, a small open model running on the district laptop, turns up to 25 seconds of speech into a fixed checklist: which symptoms, for how many days, which danger signs.
No guessingEvery "yes" must quote her own words. Anything without that evidence counts as unknown and is asked on the keypad.
Bob asks yes-or-no questions only for what her description didn't settle: danger signs such as fits, drowsiness or being unable to drink.
Fail safeSilence or a wrong key counts as "don't know", which can only make a case look more urgent, never safer.
Fast breathing in a young child is a danger sign in WHO's IMCI guidance, so the rules mark this case HIGH. She hears: go now to Rampur PHC; the doctor will call you.
Every urgent or uncertain case raises an alert that a clinician must acknowledge, by replying to an SMS or on the dashboard.
Nobody slips throughIf no one acknowledges, the alert moves to the next open clinic, then the referral hospital, then the district officer. A failed callback also alerts a person.
Clinicians report status with one word by SMS (OPEN, CLOSED, FULL) or a free missed call. When Rampur PHC texted CLOSED, the next patient was sent to Devgaon PHC, and the dashboard explains why.
Only trusted sendersMessages from unregistered numbers are ignored, so strangers can't change the routing.
Each case becomes a record under 200 bytes, encrypted and sent to Central by SMS. It holds no name and no phone number: a keyed code stands in for the patient.
Cost the system can carryIn the team's test run, about ₹0.23 per case for callbacks and SMS. The patient pays nothing.
Screens are real captures of Bob's clinic dashboard (bob/vitamin_bob/dashboard.html), running the team's 17 scripted test calls. All 17 passed, with no contract violations. This capture uses the keyword baseline and test transcripts, not live Gemma audio, with a dummy district and dummy numbers.
A keypad menu can route a call, but it has to ask every question, and it can't hear what the patient already said. Letting the patient speak first, then asking only what's missing, makes calls shorter. It also leaves far fewer cases stuck as "uncertain", and those cases otherwise fall on scarce clinicians.
| Keypad menu only | With Gemma 4 | |
|---|---|---|
| Uncertain calls, Hindi | 14% | 7% |
| Uncertain calls, Gujarati | 10% | 5% |
| Keypad questions per call, Hindi | 6.65 | 5.60 |
| Missed urgent cases | 0 | 0 |
Evaluation on 54 synthetic vignettes plus 10 team recordings (4 Oct 2026, Gemma 4 E2B on the demo laptop). On the 10 recordings, through the full audio path, tier accuracy was 100% in both languages. The model's mistakes lean towards over-triage, which is the safe direction.
| Missed call → callback → answer → hangup on a real phone (Samsung S24 Plus) | Passed |
| Prompt, keypad input and speech capture on a real call, also with the phone's Wi-Fi and data off | Passed |
| Outbound SMS delivered exactly once | Passed |
| Bob + Gemma + Stuart together: Hindi fever → Medium; Gujarati emergency → 108 + alert; clinic closed → High sent to the next clinic | Passed on recorded audio |
| A full live phone call with Gemma listening | Not yet tested |
| Incoming clinician SMS on hardware; Central over a real SMS link | Pending |
Not covered yet: real patients; clinician review of the danger-sign protocol (drafted by the team from WHO IMCI); dialects beyond the team's voices; noisy phone lines; conditions beyond fever, cough and diarrhoea. Gujarati prompts need a native speaker's check.
| Source | Used for | License |
|---|---|---|
| WHO Global Health Observatory (HWF_0001, HWF_0006, UHC_INDEX_REPORTED) | Health workforce density, service coverage | WHO terms of use |
| World Bank World Development Indicators | Population, rural population and share, internet users, mobile subscriptions, electricity access, TB incidence | CC BY 4.0 |
| Permenkes 24/2022 on Medical Records; Kemenkes (SATUSEHAT, SITB) | Indonesia's digital-health rails | Cited |
| Sahabat-AI model card (GoTo, Indosat) | Local-language models for Indonesian, Javanese, Sundanese | Cited |
| healthsites.io / OpenStreetMap via HDX | Indonesia facility map and record completeness | ODbL |
| Chaudhury, Hammer, Kremer, Muralidharan, Rogers (2004 draft; JEP 2006) | Provider absence on unannounced visits | Cited |
| Siyam et al. (2021), BMC Health Services Research 21(Suppl 1):691 | Recording share of consultation time, registers | Cited |
| WHO Workforce 2030 (2016) | 44.5 per 10,000 threshold | Cited |
| Natural Earth via world-atlas; Esri Light Gray Canvas | Country shapes; basemap | Public domain; Esri terms |
Each country's latest value since 2015 is used; years appear next to every number. The full analysis, with code, is the EDA notebook in the project repository.