SimPPL
AI4Health · ai4health.simppl.org

Sakhi: a mobile health intervention with the people who keep frontline care running.

Sakhi means "female friend" in Marathi. It is a mobile health intervention designed with India's Accredited Social Health Activists, the partner nonprofit's field staff, and expectant mothers themselves. Health information they can act on, delivered over the phones they already use.

Read the paper (MedicaLLM-Eval)
Jalgaon, Maharashtra · 2025
A user, on what the Sakhi chatbot meant for her care
Watch · Psychology of Technology Conference

Sakhi's field findings, presented in DC, November 2025.

Twelve minutes from Swapneel on how the Jalgaon RCT was designed, what the team learned about deploying conversational AI in community-health contexts, and the four lessons every health-AI team should pin to the wall.

What Sakhi does

Key features.

  1. 01Personalised health information based on the user's month of pregnancy and prior questions.
  2. 02Check-up reminders and behavioural nudges, calibrated with ASHA workers in the local context.
  3. 03Local-language audio-to-text and back, so users don't need to type Marathi or Hindi to be understood.
How it's built

Human-in-the-loop, by design.

Sakhi never replaces an ASHA worker. It triages questions, drafts responses, and routes anything ambiguous to a human reviewer who knows the community. The AI assists. Trust comes from the human.

The two RCTs

Where we've deployed and what we measured.

Jalgaon, Maharashtra · India

Maternal health literacy

In partnership with Aadhar Bahuddeshiya Sanstha. Two-year RCT improving antenatal awareness and folic-acid uptake, with downstream measures on postnatal practices.

  • User research: January – August 2024
  • Platform development: June 2023 – December 2024
  • Baseline survey: May 2025
  • Endline: October 2025
  • Presented at the Psychology of Technology Conference, Washington DC, November 2025

Preliminary results not for public distribution until paper review clears.

Bangladesh

Menstrual and reproductive health

In partnership with Spreeha Foundation. RCT with 250 local families on menstrual-health literacy. Won 2nd place at the MIT PKG IDEAS Social Innovation Challenge.

Multilingual evaluation platform

A reproductive-health Q&A benchmark, with Cohere AI.

You can't deploy a health AI system without first knowing how it answers in the actual languages and registers it's deployed in. We built a 1,000-Q&A dataset in reproductive-health areas, 150 publicly released, on the Weval community evaluations platform, with the Collective Intelligence Project. Co-authoring the technical paper with Cohere AI.

health-eval.simppl.orgQ&A Dataset
Sakhi benchmark three-channel community review pipeline: WHO ANC / India NHM / ANM Manual source corpus → Aya Expanse drafter and MedGemma validator generating 845 candidate Q&A pairs in English, Hindi, Marathi → doctor channel (n=11), ASHA-worker channel (rural India), nonprofit-staff channel → expert track (149 doctor-edited references), non-expert track (231 community-sourced), and a doctor calibration set (148 questions, 169 verdicts).
SAKHI BENCHMARK · THREE-CHANNEL COMMUNITY REVIEW PIPELINE
Production clinical onboarding

Sanjeevani AI: clinical onboarding deployed with India's largest fertility care provider.

We work with doctors to build AI-automated patient onboarding, deployed in production with India's largest fertility care provider across multiple cities. Sanjeevani is the third leg of our health stack, alongside Sakhi and Health-Eval.

About SanjeevaniHealthReview clinician console
Lessons from a two-year deployment

What we'd tell another health-AI team.

  • Design for collaboration, not replacement.

    Build workflows where AI assists, while human insight ensures relevance and accountability.

  • Do not overhype AI.

    Participants will buy the hype. You won't be able to dial it back later.

  • Conversational agents lack conversationality in medicine.

    Off-the-shelf LLMs answer the question but miss the conversation around it.

  • Lack of contextual awareness really hurts.

    Long-term memory and case-history retrieval matter more than model size.

Funder, partner, or fellow health-AI team, let's talk.

Contact