Work / Bharat-First / Janani Suraksha
No. 02 · Bharat-First · Maternal HealthLive signal · India · 2026
22,500 deaths a year. Most are preventable. India has 1.18 million ASHA workers walking village to village with a register and a pen. Janani Suraksha hands them three constant-time AI engines on the phone they already own. Predict the danger before it arrives.
70,000 entries · <5 ms · NFHS-5 calibrated
21 states · 2,823 blood banks · O(1) lookup
7,480 profiles · learned index · 2,497 params
Act I · The Problem
1.18 million workers visit every pregnant woman every month. Their training is excellent. Their pen is honest. The question they cannot answer is which of these forty mothers is the next emergency.
The right hospital has a functional OT, a blood bank with the matching group, and a specialist on duty tonight. None of that fits on a roadside signboard.
Hemoglobin trends across pregnancy. By the third trimester, the trajectory is a death sentence or a routine delivery. Today, nobody draws that line.
The tool has to run on the device the ASHA already owns. No GPU. No cloud round-trip. No assumptions about bandwidth or battery.
Act II · The Promise
Every prediction is a hash lookup. Nothing trains at runtime. The model lives in 33 megabytes baked into the container image, scales to zero, and answers before the next sentence finishes.
70,000 precomputed entries cover every realistic combination of twelve risk factors: age, parity, hemoglobin, blood pressure, gestational age, complication history, height, weight, more. A single hash returns Low, Medium, High or Critical. Every weight is cross-validated against five independent published sources.
70,000 entries <5 ms response 33 MB on disk NFHS-5 · WHO · Cochrane · Lancet · ACOG
Real facilities from data.gov.in National Hospital Directory, with geocoordinates. 2,823 blood banks. Twenty-one states and two union territories. A precomputed spatial index per capability level returns the nearest hospital that has the OT, the specialist and the matching blood. Google Maps opens with one tap for navigation.
21 states 2,823 blood banks O(1) primary lookup data.gov.in source
A two-layer MLP (5 to 64 to 32 to 1, 2,497 parameters) predicts the position in a sorted trajectory array. Constant time. The output is a sentence the ASHA can act on: "with 90 percent IFA compliance, hemoglobin improves from 7.2 to 9.8 g/dL." Inspired by Kraska et al. learned indices, applied to maternal hematology.
7,480 trajectories 2,497 parameters O(1) position lookup WHO-calibrated physiological model
Act III · The Evidence
Risk in pregnancy is not a guess. It is a literature problem. Janani Suraksha pulls from twelve risk factors, cross-validates every coefficient against five independent sources, and refuses to ship a number that no one published.
Source · 01
National Family Health Survey 2019-21. Indian baseline rates for parity, anemia prevalence, and antenatal coverage.
Source · 02
Maternal mortality framework, anemia thresholds, eclampsia management protocols.
Source · 03
Systematic review weights for IFA supplementation, blood pressure intervention, and gestational diabetes.
Source · 04
Maternal Health Series for relative risk coefficients on age, complication history, and BMI.
Source · 05
American College of Obstetricians for cross-validation against high-income clinical practice.
Source · 06
National Hospital Directory with geocoordinates, plus 2,823 blood banks. Government of India, Ministry of Health and Family Welfare.
Source · 07
Sample Registration System: maternal mortality ratio of 88 per 100,000 live births, the figure this product attacks.
Source · 08
The accepted obstetric delay model used across low and middle income countries to design interventions.
Source · 09
Practice guidelines: every prediction requires human clinical confirmation. The ASHA is in the loop. Always.
Act IV · Proof
janani-suraksha-pax2obvj3a-el.a.run.app
FastAPI on Google Cloud Run. Min instances zero, scales to thousands. ~$0/month at low traffic, no GPU at runtime, no database round-trips.
make test
Risk scoring, referral routing, anemia prediction, full API contract. Every endpoint covered. CI gates every commit.
make deploy
Infrastructure as code. Cloud Run, Artifact Registry, IAM. New region in minutes. Tear down with make destroy.
defense-in-depth
Cloud Run IAM. Sliding-window rate limiting. Pydantic strict input validation with medical ranges. CSP, CORS, security headers. Audit log per assessment. Non-root container.
/dashboard
Aggregated risk distribution, hotspot identification, intervention effectiveness. The block-level view a CMO actually opens at 9 AM.
/api/voice
The ASHA does not have to type. She speaks. Hindi, Bengali, Telugu, Marathi, Tamil, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese, English.
The Stack
I build constant-time, evidence-cited AI for the parts of India where bandwidth, devices and decisions are all scarce at once. Healthcare, agriculture, civic infrastructure. If your problem looks like that, let's talk.