Care that starts with a missed call_
Any phone. No internet. Free for the patient. Bob calls back, listens in Hindi or Gujarati, and sends you to the clinic that is actually open. Not sure? A person decides.
That is one health worker for about 270 people. Their time is the scarcest thing in the clinic, and today it goes to whoever arrives first.
Source: WHO Global Health Observatory, India 2024 (doctors, nurses and midwives); WHO SDG threshold 44.5 (2016). 270 = 10,000 ÷ 36.7.
The urgent wait with the routine.
Without a way to sort people before they travel, the clinic sees whoever came first. Somewhere in the line, a child has had a fit. Nobody at the door knows.
Source: Illustrative. First-come queues, long travel and little time per case are described in the World Bank brief, Annex A (2026).
of every consultation goes to writing in registers. A clinic can keep up to 48 of them, and the same details are written again and again.
Source: Siyam et al. (2021), BMC Health Services Research 21(Suppl 1):691, five countries.
What if the clinic’s front desk lived inside a missed call?
Not an app. Not a smartphone. The call anyone can already make, answered by a small AI that listens, sorts and routes, so the scarce clinician’s time goes to the people who need it most.
She rings once and hangs up. It costs her nothing.
- Works on any phone: no app, no data plan
- The district pays for the callback
Bob calls back, in her language.
- Hindi or Gujarati; another language is data and audio files, not code
- Every sentence is a fixed recording: the AI never speaks
She describes the problem once, in her own words.
- Gemma 4 runs on the district laptop: no internet
- A 25-second clip is understood in under 20 seconds
Source: Measured on the demo laptop (RTX 3050 Ti, 4 GB). Gemma 4 E2B, 8-bit, via llama.cpp.
Every answer must be backed by her words.
- No quote, no answer: it is asked on the keypad instead
- What she already said is not asked again
Written rules decide. Not sure? Ask a person.
- Five levels of urgency, each with its reasons
- Anything that could hide an emergency goes to a clinician
The nearest clinic that is actually open.
- Clinics report OPEN, CLOSED or FULL by SMS or a free missed call
- Emergencies always go to 108
Urgent cases land on a clinician’s phone.
- One reply acknowledges it
- No answer in 3 minutes: it moves to the next clinic, then the district
The record writes itself.
- Every case: urgency, reasons, rules and model version
- Sent to the state over SMS, 200 bytes, no names or numbers
Does it work? We measured it.
urgent cases missed, in every test set and both languages
the “not sure” rate when Gemma listens, compared with keyword matching
paid by the patient; about ₹0.30 per case for the district (estimate)
| CASES | NOT SURE · KEYWORDS | NOT SURE · GEMMA |
|---|---|---|
| Hindi, 43 cases | 14% | 7% |
| Gujarati, 21 cases | 10% | 5% |
| 10 real voices | – | 0% |
Source: Bob’s own evaluation: 54 synthetic cases written by the team and 10 recordings of team members’ voices; not real patients. Labels not yet reviewed by a clinician.
Watch a real call.
Give a missed call to the district line and follow it: received, called back, connected, in conversation, and the record Bob keeps at the end. Everything runs on one laptop.
The phone line, Gemma and every patient’s data stay on one laptop at the district hub, with no internet. This public copy of the site cannot reach it, by design. Run it yourself to watch calls live, or see the demo video.
How to run it →- 1Missed call received
- 2Queued for callback
- 3Bob rings back
- 4In conversation
- 5Result recorded
Run it on one laptop.
Everything runs offline on a Windows laptop. A built-in simulator plays the patients, so you can watch full calls in Hindi and Gujarati without a phone.
git clone https://github.com/Meet2304/vitamin-bob
# + llama.cpp and Gemma 4 E2B (see README)cd bob
tools\start_model.cmd
.venv\Scripts\python -m vitamin_bob.server.venv\Scripts\python -m fake_stuart run --pace 1.5
# open http://127.0.0.1:8100/dashboardLocalizing AI means meeting people on the phone they have, in the language they think in, and knowing when to step aside.
A small model on one laptop is enough to listen and sort. The decisions that matter stay with people.
Meet the four of us who built it →Built for the World Bank × Hack-Nation Small AI for Development hackathon, health track, October 2026. Sources: World Bank concept note (2026), WHO Global Health Observatory, Siyam et al. (2021). Not a diagnostic tool; the protocol is drafted from WHO IMCI and not yet clinician-reviewed. Demo clinics and numbers are placeholders.