Women's health · Voice AI

Vaani Sakhi वाणी सखी

A clinically grounded voice-AI agent for women's health, starting with postmenopausal care. A woman, or her ASHA worker, phones it and talks in Hindi or English.

In brief

Indian women reach menopause around 46,11 Indian Menopause Society survey, 2016, and Babbar et al., Sci Rep 2024, using national survey data. ↓ about five years earlier than women in the West. Heart, bone, urinary and breast risks all rise in the years that follow, yet most women never talk to anyone about it. Vaani Sakhi is a phone voice agent that answers their questions. It never answers from the AI model's memory. Clinical statements come only from guidelines that clinicians have signed off,22 Indian Menopause Society 2026, The Menopause Society, NICE NG23 and India's programme guidelines for non-communicable diseases. ↓ risk numbers come only from validated calculators the agent calls as tools,33 The WHO South-Asia heart risk chart, FRAX for fracture risk, and the Menopause Rating Scale. ↓ and every warning sign goes to a human.

It builds on Palli Sahayak, our earlier voice helpline for palliative care, and reuses its engine. It works in two modes: a woman can call it herself, or an ASHA worker44 Accredited Social Health Activists: community health workers under India's National Health Mission, about one for every 1,000 rural people. ↓ can call it during a home visit. The longer-term aspiration is a sakhi, a trusted friend, that gets to know each woman over many calls.

~46 yrs
Average age at menopause in India, about five years earlier than in Western women
IMS survey 20161
~2%
Breast-cancer screening coverage among Indian women
South Asian J Cancer 20175
7.8/10
Palli Sahayak expert score over 160 reviews. 59% endorsed it without reservation
Earlier helpline evaluation
≥85%
Target: answers to test cases that clinicians judge safe
Vaani Sakhi safety bar

The problem

Five organ systems change at once

During the menopause transition, LDL ("bad") cholesterol and arterial stiffness rise,66 El Khoudary et al., Circulation 2020: the menopause transition and heart disease risk in women. ↓ bone loss speeds up, and genitourinary symptoms such as vaginal dryness and urinary urgency begin.77 Together called the genitourinary syndrome of menopause. Portman and Gass, 2014; AUA/SUFU/AUGS guideline, 2025. ↓ Women also face decisions about hormone therapy and breast cancer. Each of these has its own guideline, written by a different specialist society.

Separate guidelines, one woman

HeartLDL cholesterol and arterial stiffness riseWHO South-Asia CVD risk chart
BoneBone loss speeds up, fracture risk climbsFRAX-India, IOF guidance
GenitourinaryDryness, pain, urinary urgencyAUA/SUFU/AUGS 2025
Breast and cervixScreening is rare, decisions are hardICMR screening, family-history rule
Symptoms and mindHot flushes, sleep, moodMenopause Rating Scale, crisis flags
One plan she can actually followcombining the advice, and fitting it to the conditions she lives and works in
  • work
  • diet
  • family decision-making
  • money
  • language and literacy
  • privacy
Hormone therapyIt eases hot flushes and protects bone, but breast and blood-clot risk limit who should take it. Vaani Sakhi explains the options and always refers the decision to a clinician.
ExerciseWeight-bearing exercise helps bone, but painful knees or a heart condition change which exercise is safe for her.
The core challenge is whole-body advice. Each organ system's guideline can recommend things that conflict with another's. The two tensions shown are illustrations. Combining guideline-backed advice into one personal plan is the hard problem we are working on. We do not claim to have solved it.

What Palli Sahayak taught us

Palli Sahayak, our earlier helpline for palliative care, has the engine Vaani Sakhi reuses: hybrid search over guidelines, speech recognition in all 22 scheduled Indian languages, a five-pillar safety framework and a warm handoff to a human. Its expert evaluation showed where it failed, and Vaani Sakhi is designed around those failures.

Most failures came in difficult escalation cases
Escalation rules and thresholds co-written and signed off by clinicians
Built for palliative care
The same engine taken into midlife women's health, with fixed, validated risk calculators
Experts who had not been calibrated disagreed sharply
Calibrated raters, two ratings per answer, a senior adjudicator
Evaluated by experts on written cases
Also tested with real users, under clinician supervision

Architecture

How a call flows through the system

The caller uses an ordinary phone line, so a basic feature phone without internet works. Personal identifiers are stripped in India before anything reaches a cloud model. Google's Gemini Live API handles real-time speech to speech for the proof of concept.88 Gemini Live streams audio both ways and can call functions, which is how the four tools plug in. ↓ If it is down, slow or weak in a dialect, the call switches to the Palli Sahayak speech pipeline from Sarvam AI and an open-weight model hosted in India.99 Sarvam AI builds Indian-language speech models. An open-weight model publishes its parameters, so it can run on our own servers. ↓ Whichever model answers, it must go through the same four tools.

System architecture

Woman or ASHAany phone, Hindi or English
Phone line and IVRno internet needed
De-identificationnames and numbers strippedIndia
Gemini Live APIprimary: real-time speech to speechcloud
if down, slow or weak in a dialect
Sarvam speech + open-weight modelfallbackIndia
Future: her health profile, kept with consent, gives the tools contextIndia

Same four tools, whichever model answers

retrieve_guideline
Knowledge basesigned-off guidelines + knowledge graph
compute_risk
Validated calculatorsWHO CVD · FRAX-India · MRS
check_red_flags
Red-flag rulesCBAC + crisis signs
handoff
Clinician or ASHAcall-back + SMS summary
Audit log: every answer stored with the model, knowledge-base and pathway versions that produced itIndia
One engine, two providers, four tools. The model is a conversational layer. Knowledge, numbers and escalation all live in parts we control and can test. Every answer is logged with the versions of the model, knowledge base and pathways that produced it. The dashed profile store is a future aspiration.

The rule the agent follows on every turn

Drawing diagram…

Swipe sideways to see the whole flowchart.

Safety checks come first, then numbers, then guideline text. A simplified view of the decision logic. Red flags short-circuit everything else. Questions about hormone therapy or prescriptions always end in a referral, and the agent never prescribes. A family history of breast cancer triggers referral for risk assessment rather than a numeric score, because no breast-cancer score is validated in Indian women.

Usage workflow

A call, step by step

Here is a made-up example in woman mode. A 51-year-old woman calls from a feature phone. She has night sweats, her knees hurt, and her mother broke a hip.

Drawing diagram…
  1. She says"I wake up drenched at night, my knees hurt, and my mother broke her hip."
  2. Toolcheck_red_flags: none found.
  3. AgentAsks the Menopause Rating Scale questions, then compute_risk (MRS) returns her symptom score and severity band.
  4. AgentAsks her age, weight, height and past fractures, then compute_risk (FRAX-India) returns her 10-year fracture risk.
  5. Toolretrieve_guideline on hot flushes and bone health returns signed-off passages with source and date.
  6. AgentGives self-care advice in her language, with sources.
  7. She asks"Should I take hormone tablets?"
  8. AgentExplains the options and never prescribes.
  9. HandoffA clinician or ASHA is asked to call her back, and she gets an SMS summary to take to the doctor.

If she had mentioned bleeding after menopause, check_red_flags would fire and the call would go straight to handoff.

The model talks, the tools decide. Every number she hears comes from a calculator, every clinical statement from a signed-off guideline passage with its source, and the hormone-therapy question goes to a clinician with an SMS summary she can take along.

Who it serves

Two modes, one engine

Woman mode

For a midlife woman, on her own phone

  1. She calls and describes what is bothering her, in her own words.
  2. The agent screens for red flags, then scores her symptoms and risks with the calculators.
  3. It explains self-care from signed-off guidelines, in plain language.
  4. It helps her prepare for a doctor's visit: what to mention, what to ask.
  5. Anything that needs a clinician becomes a call-back request and an SMS summary.

ASHA assist mode

For an ASHA worker, during a home visit

  1. The ASHA fills in the CBAC checklist with the woman, as she already does.
  2. She calls the agent for follow-up on the items the checklist raised.
  3. The agent checks the answers against CBAC red flags, such as bleeding after menopause or a breast lump.
  4. It suggests what the ASHA can explain on the spot, and what needs a referral.
  5. Red flags go to a clinician, with an SMS summary to the ASHA.

Both modes share the same knowledge base, tools, red-flag rules and audit log.

Tools instead of recall

The four tools

Gemini Live can call functions: the model runs a piece of our own software and uses its result. That is how the agent stays grounded.

retrieve_guideline

Hybrid search, combining keywords and meaning, over the curated guidelines. Returns passages with source, date and evidence grade.

  • IMS 2020 and 2026, The Menopause Society 2022 and 2023, NICE NG23
  • AUA/SUFU/AUGS 2025, IOF, ICMR screening, NP-NCD and CBAC
  • Rewritten as structured text and a knowledge graph
compute_risk

Three simple, validated calculators, nothing else.

  • WHO South-Asia non-lab CVD chart: 10-year heart attack or stroke risk
  • FRAX-India: 10-year major fracture risk
  • Menopause Rating Scale: 11-question symptom score

No numeric breast-cancer score. Family history triggers a referral instead.

check_red_flags

Looks for warning signs that need prompt medical attention.

  • CBAC items, such as postmenopausal bleeding or a breast lump
  • Warning signs of heart disease in women
  • Signs of a mental-health crisis
handoff

Asks a clinician or the woman's ASHA to call her back, and sends her an SMS summary.

  • Every red flag
  • Every hormone-therapy or prescription question
  • Anything the guidelines do not cover

Safety and data

What stays in India, what reaches the cloud

Stays in India

  • Name, phone number and other identifiers
  • De-identification step
  • Fallback speech pipeline and open-weight model
  • Audit log with model, knowledge-base and pathway versions
  • Her health profile, in future, with consent
de-identified turns only

Cloud model

  • Gemini Live, for the proof of concept
  • Sees conversation turns with identifiers removed
  • Nothing she shares is used to train models
A provider-agnostic design. To scale up and keep costs down, we will also test open models that can be hosted in India, such as MedGemma, on the same vignettes and safety checks as Gemini. We will switch to a locally hosted model only if it does well in that testing. The design follows Indian and WHO guidance on AI in health, and India's data protection law.1010 ICMR ethical guidelines for AI in healthcare, 2023; WHO guidance on large multimodal models, 2024; Digital Personal Data Protection Act, 2023. ↓

Testing

How we test it before real users

  1. 1. Tool checksEvery calculator must match its official version on hundreds of made-up patients, and the speech layer must pick out the right values.
  2. 2. Test casesBilingual written cases rated by calibrated clinicians, two ratings each, with a senior clinician settling disagreements.
  3. 3. Safety gateReal users only if the agent clears the safety bar.
  4. 4. Supervised useWomen and ASHAs use it alongside usual care, never instead of it, with a clinician reviewing calls.
≥85%
of test-case answers judged safe by clinicians: the headline target
59%Earlier
≥85%Target
<1%
critical-error rate
≥95%
red-flag sensitivity, and no emergency left without escalation
κ ≥ 0.7
agreement with clinicians on risk tier (Cohen's kappa)
≥95%
of clinical claims traceable to a source
SUS ≥ 70
System Usability Scale score with real users
Per language
word error rate and response delay, reported separately
Benchmarks. The agent is compared with the Palli Sahayak engine, the same Gemini model without tools or retrieval, and a leading general-purpose LLM. Off-the-shelf LLMs answer only 30–70% of menopause questions fully correctly1111 A 2025 study in Menopause testing general chatbots on menopause questions. ↓, while tool-calling agents reach about 88% on clinical calculations1212 AgentMD, Nature Communications 2025: an agent that picks and runs clinical calculators. ↓. Health outcomes such as screening uptake need a larger trial, and we do not claim them.

Where it's going

A sakhi that gets to know her

Sakhi is the Hindi word for a woman's trusted friend. Menopause shows up differently in every woman, so generic advice often misses. With her consent, the agent could keep a profile of what she chooses to share across calls and start each call from it. We are designing this profile and its privacy rules with women, ASHAs and clinicians, and it is an aspiration rather than a finished feature.

Her bodyAge at menopause or hysterectomy, conditions, medicines, risk-tool results
Her symptoms over timeMenopause Rating Scale scores from call to call
Her lifeLanguage, literacy, diet, work, family decision-making, privacy needs
Changes how advice is explained, never what guidelines allowStored in IndiaShe can hear or delete it anytimeNever used to train models

Related

Go deeper

References

  1. Indian Menopause Society survey, 2016; Babbar et al., Sci Rep 2024. ↩
  2. IMS 2020; IMS 2026; The Menopause Society 2022; 2023; NICE NG23; NP-NCD operational guidelines and CBAC, 2023. ↩
  3. Kaptoge et al. WHO CVD risk charts. Lancet Glob Health 2019; FRAX; Heinemann et al. Menopause Rating Scale. 2004. ↩
  4. National Health Mission: ASHA. ↩
  5. Breast-screening coverage in India. South Asian J Cancer 2017. ↩
  6. El Khoudary et al. Menopause transition and cardiovascular disease risk. Circulation 2020. ↩
  7. Portman and Gass. Genitourinary syndrome of menopause. Menopause 2014; AUA/SUFU/AUGS guideline, 2025. ↩
  8. Gemini Live API; Gemini function calling. ↩
  9. Sarvam AI speech models; MedGemma. ↩
  10. ICMR ethical guidelines for AI in healthcare, 2023; WHO guidance on large multimodal models, 2024; DPDP Act, 2023. ↩
  11. Accuracy of large language models on menopause questions. Menopause 2025. ↩
  12. AgentMD. Nat Commun 2025. ↩