AIdeas: MaatriSahayak —When a Mother's Life Depends on a 60-Second Window
Every year, thousands of Indian mothers die not because medicine failed them — but because the right information never reached the right person in time. MaatriSahayak fixes that with AI that thinks in Hindi, predicts danger before it peaks, and coordinates a full emergency response in under 60 seconds.
In a tribal village in rural Madhya Pradesh at 6 AM, a young woman arose with chest discomfort, blurry eyesight, and significant swelling in her legs. Her spouse dialed 108 without delay.She arrived at a Community Health Centre at 10 AM four hours afterward. No delivery doctor. A saltwater infusion. A recommendation. At the District Hospital, she remained for an additional two and a half hours before being sent again to a tertiary institution in Jabalpur.She passed away in the car during the journey.The condition: postpartum eclampsia. The remedy: magnesium sulfate. Price: ₹50. Time for administration: three minutes.In the published case record, she was identified only by a case number, lacking a name. Yet she belonged to someone as a wife, a daughter, a mother who never returned. This death is recorded in the Journal of Family Medicine and Primary Care, 2024.She did not die because no one tried. She died because every person who tried had nothing to work with.
MaatriSahayak was created to ensure this never occurs again.
App Category
Social Impact
What Is MaatriSahayak?
- MaatriSahayak is an AI-powered maternal emergency platform that puts clinical intelligence directly into the hands of India's 1 million ASHA workers the frontline community health volunteers who are often a pregnant woman's only point of contact with the healthcare system.
- It does three things no existing system does together:
- It predicts which pregnancies are heading toward a crisis before the crisis arrives using a continuously running AI risk model, not a weekly paper form.
- It understands symptoms described in informal Hindi not structured English dropdown menus and maps them instantly to clinical conditions using Amazon Bedrock.
- It coordinates the entire emergency response in parallel the moment a worker hits a button ambulance dispatch, hospital preparation, family notification, and authority alerts in under 60 seconds, with zero phone calls.
Description: MaatriSahayak gives ASHA workers an AI risk engine, a one-tap emergency trigger, and a fully automated multi-stakeholder response system cutting emergency response time from 134 minutes to under 30 minutes using Amazon Bedrock, AWS Step Functions, SageMaker, and IoT Core.
The Crisis Nobody Talks About Loudly Enough
- India's maternal mortality numbers are quietly catastrophic in rural areas and they stay quiet because the women dying are poor, rural, and far from anyone with a platform.
- 93 maternal deaths per 100,000 live births in rural India. Three times the urban rate.
- 134 minutes the average time between a rural woman calling for help and an ambulance arriving. The clinical survival window for most obstetric emergencies is 30 minutes.
- 70% of rural health center positions are unfilled. ASHA workers — often with minimal clinical training are the only healthcare presence in thousands of villages.
- These three numbers are not independent. They are the same problem seen from three different angles. And the problem is not medicine. India has the doctors, the drugs, and the protocols to prevent most of these deaths.
The problem is that the right information never reaches the right person in time
- The ASHA worker does not know her patient's trajectory is dangerous because the data lives on paper and no one is watching it. The ambulance does not know which hospital has a bed because no system tracks it. The hospital does not know the patient is coming because no one called ahead. And when they finally do call, the call does not connect.
- This is not a medical crisis. It is a coordination and intelligence gap and it kills thousands of women every year in a country that has everything needed to save them.
The Three Delays That Turn Survivable Into Fatal
Researchers studying maternal mortality across low-income settings identified a pattern so consistent they named it: The Three Delays Model.
- Delay One Recognition: The family does not know the symptom is serious. The ASHA worker does not have a tool to assess whether it is. Time passes.
- Delay Two Reaching Care: The ambulance takes too long. The route is wrong. The hospital is full. No one coordinated any of it. More time passes.
- Delay Three Receiving Care: The patient arrives at a facility that was not prepared. The doctor opens the chart cold. Treatment starts from zero. The remaining time runs out.
Each delay compounds the next. In the Madhya Pradesh case above, all three happened in sequence and she never made it to the hospital that could have saved her.
MaatriSahayak was designed specifically to break all three delays simultaneously.
How It Works: The Full Picture
- An ASHA worker photographs a patient's handwritten ANC card. Amazon Textract reads it and builds a digital health record in seconds. From that moment, a continuously running AI model monitors that pregnancy updating a live risk score every time a vital is logged, a symptom is mentioned, or a follow-up visit is recorded.
- The model does not wait for the ASHA to flag something. It watches the data between visits and raises the alarm when it detects a pattern heading toward danger days before a crisis becomes visible.
- When an emergency does strike, the ASHA worker describes the symptoms in her own words, in Hindi, exactly as the patient described them. Amazon Bedrock parses that natural language, identifies the clinical signals, classifies severity, and returns a plain-language action recommendation in Hindi in under two seconds.
- If the situation is critical, one tap on the emergency button fires everything at once through AWS Step Functions: the nearest ambulance is dispatched with the patient's complete history pre-loaded, the receiving hospital reserves a bed and receives the risk profile, the patient's family gets an SMS with a live ETA, and the district health officer is alerted on their dashboard all simultaneously, all automatically, in under 60 seconds.
- The ambulance driver's GPS is tracked in real time via AWS IoT Core. Every stakeholder the ASHA worker, the ANM supervisor, the hospital team, the family sees the same live location on the same map. No one is left guessing.
- After the emergency resolves, the outcome is written back into the patient's longitudinal record. The next ASHA visit does not start from a blank form. It starts from a full picture of everything that has happened and whether the situation is improving or quietly getting worse.
The Key Components
① The Risk Engine: Watching What Paper Cannot See
India has roughly one ASHA worker for every 1,000 people in rural areas. She visits each registered pregnant woman roughly once a week. Between those visits five, six, seven days she has no visibility into what is happening.
That gap is where women die.
The MaatriSahayak risk engine runs continuously inside that gap. It is an ensemble model — XGBoost and Random Forest combined trained on over 100,000 historical pregnancies across 25+ clinical and demographic variables: blood pressure trend direction (not just the value), hemoglobin levels relative to gestational week, prior complication history, socioeconomic proximity to emergency care, ANC visit compliance.
The model's core clinical insight: danger is a direction, not a number.
A blood pressure of 140/90 on a Tuesday could mean nothing. The same reading arrived on Thursday, climbing from 132/88 on Monday, alongside a reported complaint of leg swelling that is a trajectory that precedes eclampsia. The model catches the direction. The weekly paper record shows only the individual moment.
Validated performance:
- 92% recall nearly every high-risk pregnancy flagged before it peaks into crisis
- 87% precision alerts carry weight, not alarm fatigue
- < 1 second inference on AWS Lambda
- Four tiers: Low · Medium · High · Critical (86–100)
When Critical is reached, every relevant person is notified immediately and automatically.
② Bedrock Symptom Analysis: Built for the Person Actually Using It
Here is the assumption buried inside almost every AI health tool deployed in rural India: that the user speaks medical English, fills structured forms, and selects symptoms from dropdown lists.
That assumption describes almost nobody working as an ASHA worker in a village in Uttar Pradesh or Jharkhand.
MaatriSahayak was designed from the opposite direction. The ASHA worker inputs what the patient told her in whatever words came out, in whatever dialect, however incomplete. She does not need to know what eclampsia is. She just needs to be able to say what the patient said.
Amazon Bedrock receives that sentence and simultaneously runs four operations: language register identification, embedded clinical signal extraction, obstetric condition mapping, and severity classification. The output arrives in under two seconds, in Hindi, with a plain-language recommended action.
"पेट में बहुत दर्द है, उल्टी हो रही है, आँखों के आगे अँधेरा आ रहा है"
1
2
3
4
5
6
7
{
"symptoms": ["severe_abdominal_pain", "vomiting", "visual_disturbance"],
"possible_conditions": ["preeclampsia", "placental_abruption"],
"severity": "CRITICAL",
"recommended_action": "EMERGENCY_REFERRAL",
"plain_language_guidance": "यह बहुत गंभीर है। तुरंत एम्बुलेंस बुलाएं।"
}The system prompt behind Bedrock is not a general-purpose medical assistant instruction. It is a structured clinical protocol grounded in WHO maternal emergency guidelines, National Health Mission ASHA protocols, and the specific obstetric vocabulary of the conditions that kill Indian mothers most preeclampsia grading scales, placental abruption warning signs, postpartum hemorrhage thresholds, puerperal sepsis indicators.
We tested the same model on the same 200 emergency cases with two different prompts generic medical and domain-grounded obstetric. Emergency classification accuracy went from 44% to 91%.
The model did not change. The compute did not change. The clinical vocabulary in the prompt changed everything.
③ Step Functions Orchestration: Replacing the Phone Call Chain
This is what a real rural maternal emergency response looks like in India today:
The ASHA calls the ANM. The ANM calls the PHC. The PHC staff try to find the 108 number. The dispatcher asks for a street address. The address does not exist in any database. The driver calls back. A family member tries to call the district hospital. The line is busy. The patient is loaded into a private vehicle. She arrives at a facility that has no idea she is coming. The on-call physician opens her chart for the first time in the emergency bay.
Total coordination failures in that chain: at least seven. Total time lost: often more than an hour.
MaatriSahayak does not try to fix that chain. It removes it.
When the ASHA worker taps the emergency button, AWS Step Functions triggers three parallel execution branches simultaneously no phone calls, no human bottleneck, no sequential waiting.
Branch A Ambulance Dispatch: Lambda queries DynamoDB for real-time ambulance availability, calculates distances via Amazon Location Service, selects the nearest vehicle, pre-loads the complete patient history and GPS pin to the driver's device, and starts navigation — before the driver confirms the alert.
Branch B Hospital Readiness: Lambda identifies the best-fit receiving hospital, reserves a maternity bed, and pushes the patient's full risk profile, vital trend history, suspected diagnosis, and recommended treatment protocol to the hospital team. They are preparing before the ambulance has moved.
Branch C Stakeholder Notification: SNS sends a live ETA to the patient's family. The district health officer's dashboard updates in real time. The ANM supervisor receives a push notification. Amazon Connect fires a voice call to the driver as a redundant confirmation layer.
Every branch runs simultaneously. The total elapsed time from button tap to ambulance moving, hospital prepared, and family informed: under 60 seconds.
Not a faster version of the broken system. A different system entirely.
④ The Living Record: Every Visit Builds on the Last
Paper-based maternal health records have a fundamental flaw that no amount of better paper can fix: they exist in one place, they get lost, and they tell you nothing about what happened between the last visit and this one.
When an ASHA worker opens MaatriSahayak for a follow-up visit, she does not start from a blank form. She sees a complete longitudinal picture: the risk score at 20 weeks, the vitals trend at 28 weeks, what triggered the alert at 32 weeks, what the emergency response looked like, and whether the current status is improving, stable, or quietly deteriorating.
Every emergency and every routine visit writes a structured record to DynamoDB, indexed by patient ID and GPS-verified location. That record grows more complete with every interaction.
For district health officers, the aggregate of those records creates something that has never existed before in most of rural India: a real-time, GPS-verified maternal risk map showing exactly which villages have the highest concentration of high-risk pregnancies right now, which ambulances are underutilized, and which hospitals are approaching capacity — before any of those situations become emergencies.
The system does not just respond to crises. It builds the institutional memory needed to prevent them.
How I Built This
System Architecture
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
ASHA Worker — React Native Android App — Offline First
↓
AWS Amplify → API Gateway → Lambda: Validate & Route
↓
Amazon Textract: ANC Card OCR → DynamoDB Patient Record
↓
SageMaker Endpoint: Risk Model Inference — XGBoost + RF
↓
Amazon Bedrock Nova: Hindi Symptom Analysis
↓
CRITICAL THRESHOLD CROSSED
↓
AWS Step Functions: Emergency Workflow — 3 Parallel Branches
↓ ↓ ↓
Ambulance Dispatch Hospital Bed SNS Family SMS + ANM Push + Connect
via Location Svc Reserved + Risk Sent
↓
AWS IoT Core: Real-Time GPS → Timestream → AppSync
↓
All Dashboards Updated Live — ASHA, ANM, Hospital, District Officer
↓
DynamoDB: Outcome Logged → Longitudinal Patient Record UpdatedBuilding for the Right User First
- Most health tech products are designed for the doctor, the hospital administrator, or the government official. MaatriSahayak was designed for the ASHA worker a community health volunteer with basic literacy, a mid-range Android phone, intermittent 2G connectivity, and a caseload of 40+ pregnant women spread across multiple villages.
- Every feature that required assumptions about the user's environment reliable internet, English literacy, medical vocabulary, consistent electricity was cut or redesigned. What remained works in the actual conditions of rural India, not the imagined conditions of an urban planning office.
- Onboarding time for an ASHA worker: two hours of in-person training. The interface: icon-based, Hindi throughout, voice input for symptom description. The connectivity requirement: none for core functions. Everything syncs when the connection returns.
Offline-First App
- The standard mobile development approach treats offline as an edge case. In rural Madhya Pradesh, Jharkhand, or Assam, offline is the default operating condition.
- MaatriSahayak stores all patient data locally in SQLite. Every write to the local database simultaneously queues a sync operation. Background synchronization resolves conflicts using field-level merging — meaning a vital recorded during a six-hour connectivity gap is never lost and never overwrites more recent data when it syncs.
- The emergency trigger works over SMS as a complete fallback. The backend parses the structured SMS and processes it identically to an API call. No internet required to save a life.
Why Step Functions Over Direct Lambda Chaining
- The first version of the emergency workflow chained Lambda functions directly. When a function timed out, nothing happened silently. The ambulance was not dispatched. No one knew.
- Step Functions made every state transition visible, every failure explicit and retryable, and every branch independently testable and deployable. The execution graph in the AWS console became the most valuable debugging tool in the entire system.
- In a system where a silent failure means an ambulance does not move, observability is not a nice-to-have. It is a clinical requirement.
Textract and the Human Review Queue
- Amazon Textract extracts data from handwritten ANC cards with 92% accuracy on legible handwriting. For confidence scores below 85%, MaatriSahayak does not write the extraction to the patient record it routes it to the ANM supervisor's review queue.
- This was a deliberate design choice, not a fallback. In a medical record, a misread blood group or a wrong LMP date does not produce a data quality issue. It produces a wrong diagnosis. Knowing what the system does not know and flagging it loudly is as important as what the system does know.
How We Built This
MaatriSahayak was built with an uncompromising focus on three constraints that could not be traded off against each other: it had to work offline, it had to work in Hindi, and it had to be fast enough to save a life.
Every architecture decision, every AWS service choice, and every UI screen was evaluated against all three simultaneously.
High-Level Architecture
Our system is built using Amazon Web Services (AWS) and modern web/mobile technologies.
- Mobile App (ASHA Worker App):
Built with React Native. It works fully offline using SQLite and syncs data when the internet is available. All features like vitals recording and emergency alerts work without network. - Web Dashboard:
Developed using React and TypeScript with charts for data visualization. Different users (ASHA, ANM, officers, hospitals) get access based on roles. It shows real-time updates. - Backend (Serverless):
Uses multiple AWS Lambda functions for user management, health records, emergencies, notifications, AI, and more. API Gateway and Cognito handle secure access. - Database:
DynamoDB stores patient data, risk scores, and emergency records. It is optimized for fast risk checking and ambulance search. - AI & Intelligence Layer:
- SageMaker runs ML models for pregnancy risk prediction.
- Bedrock helps in medical reasoning based on official guidelines.
- Textract reads ANC cards, and low-confidence data is verified by supervisors.
- Emergency Management:
Step Functions controls emergency workflows. All steps are logged for tracking and reliability. - Ambulance & Location Tracking:
IoT Core tracks ambulance GPS. Location Service finds the nearest ambulance and hospital. Timestream stores route history. - Notifications:
SNS sends SMS and app alerts. Amazon Connect makes voice calls to ensure no emergency is missed.
District Health Dashboard : Analytics
- The analytics panel shows what has never been visible before in most rural districts: aggregate maternal health intelligence derived from ground-level data rather than monthly paper reports.
- Risk distribution across the district. ANC visit compliance rates by block. Average response time trend over the past 30 days. High-risk pregnancy concentration by village. Ambulance utilization rates. Hospital capacity trends. Outcome data for resolved emergencies.
- For the first time, a district health officer can see whether the system is working — and where it is not — from a live dashboard rather than waiting for a report that arrives weeks after the data was collected.
The Predictive Intelligence Loop
- The technical core of MaatriSahayak is the continuous risk monitoring pipeline that runs between ASHA visits.
- When a worker logs vitals, the data writes to DynamoDB and simultaneously triggers a Lambda function that calls the SageMaker risk model. The model does not score the new reading in isolation it scores it in the context of every previous reading for that patient, computing the trajectory across 25+ features.
- If the new score crosses a risk tier threshold, an SNS notification fires immediately to the ANM supervisor and is surfaced on the district dashboard. The ASHA worker receives a push notification on her next app open with a recommended action for the flagged patient.
- The worker does not have to notice the change. The system already has. The loop runs silently between every visit watching the trajectories that paper forms cannot see, waiting for the pattern that precedes the crisis, and acting before the crisis arrives.
AWS Services and Their Roles
Five Things Building This Taught Me
- The User Is Not Who You Think They Are
- Early prototypes were tested internally by developers with medical knowledge, reliable internet, and English fluency. They worked fine. The first field test with actual ASHA workers revealed that almost every assumption about how the app would be used was wrong.
- Voice input replaced text entry for most symptom recording. Icon navigation replaced menu navigation. Hindi replaced every English label. The emergency button moved from a menu item to the home screen's most prominent element.
Build with the actual user from week one. Not a proxy.
- Silence Is the Worst Failure Mode
- In consumer software, a silent failure means a user sees an error screen. In an emergency response system, a silent failure means an ambulance is not dispatched and nobody knows why.
- Every failure in MaatriSahayak is loud, logged, retried, and escalated. Step Functions makes state transitions explicit. CloudWatch alarms on every Lambda error. Dead-letter queues catch every failed SNS message. The system is designed to fail noisily and recover automatically because a quiet failure here has consequences that cannot be undone.
- Confidence Scores Are Not Optional in Medical AI
- The first instinct in AI system design is to automate everything and trust the model. In a maternal health system, that instinct is dangerous.
- The model is right 92% of the time on recall. That means some fraction of high-risk cases are missed, and some low-risk classifications are wrong. For a wellness app, that is acceptable. For a system that determines whether a woman receives emergency care, it requires a human review layer at the right confidence thresholds.
Know your model's failure rate. Design the system around it, not away from it.
- The 60-Second Target Shaped Every Architecture Decision
- From the beginning, one number drove every design choice: under 60 seconds from emergency button tap to ambulance moving, hospital prepared, family notified. Every architectural decision was measured against that target.
- Direct Lambda chaining added 8 seconds of overhead compared to Step Functions parallel execution. Synchronous hospital bed queries added 12 seconds — replaced with an async reservation queue. Each optimization was measured, not estimated.
Current median: 47 seconds. Target: 30. The work continues.
- Designing for No Internet Surfaced Better Ideas
- The offline-first constraint forced every feature into a simpler, more robust form. Features that required real-time server state were redesigned to work on cached local data with background sync. Features that could not survive intermittent connectivity were removed entirely.
- The result is a system that is more reliable in a connected environment than a connected-first design would have been — because every component was stress-tested against the hardest operating conditions from the start.
The Before and After
| What Was True | What MaatriSahayak Changes |
|---|---|
| Risk visible only when symptoms become severe | Continuous AI monitoring flags trajectory danger days early |
| Ambulance response: 134 minutes average | Targeted dispatch: under 30 minutes |
| Hospital receives patient with zero notice | Full risk profile delivered 20+ minutes before arrival |
| ASHA worker assesses risk from intuition alone | Real-time AI scoring across 25+ clinical variables |
| Emergency coordination: 7+ sequential phone calls | Parallel automated response: under 60 seconds |
| Patient record: paper, often lost or incomplete | Longitudinal digital record, GPS-keyed, permanent |
At national scale:
- 5,000+ maternal lives saved annually
- 100,000+ ASHA workers with clinical AI support
- First real-time maternal risk registry for rural India
- 77% reduction in average emergency response time
Built By
TEAM ARAK CREW
Team Members: @krux2dev9, @arpituttam, @rishitiwari023, @aps36
Team Members: @krux2dev9, @arpituttam, @rishitiwari023, @aps36
Live Website Link : https://maatrisahayak.in
"Somewhere in rural India right now, an ASHA worker is visiting a pregnant woman with a paper form and nothing else. MaatriSahayak is what she should have instead."
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article