Sakhi: BestFriend

Inspiration

For women and girls in remote and rural areas, a simple health question can be blocked by distance, cost, and stigma all at once. Many never reach a clinic in time, and the NGOs trying to help are flying blind about where need is actually concentrated. We wanted one free, confidential phone number that meets a woman in her own language and, at the same time, tells providers exactly where to show up. Sakhi (meaning "friend") is that bridge between a quiet call for help and real care on the ground.

What it does

Sakhi is a voice-first healthcare helpline with two connected sides:

• For callers: A woman calls and speaks naturally in her own language. A Grok voice agent triages with care, builds a structured clinical note, assesses severity, and gives a ranked, escalating set of recommendations — first-aid guidance, a referral to the nearest capable clinic, and, with consent, a spot on an NGO outreach waitlist.
• For NGOs: An impact dashboard shows anonymized, aggregated demand on a live map, highlighting under-served areas as recommended camp locations. NGOs filter by need, request a camp (services, date, expected reach), and activate it.
• The loop: Activating a camp turns that area into "covered" and automatically notifies waitlisted callers nearby — so a single call becomes real care, and that care feeds back into what the helpline can offer the next person.
• Full traceability: Every AI tool call and recommendation is an append-only record, viewable as a care journey, formatted recommendations, and a complete audit timeline.

How we built it

• Next.js (App Router) + React on Vercel for the app and API routes.
• xAI Grok & Grok Voice Agent API for multilingual speech-to-speech intake with tool calling, plus structured note + severity generation. The browser mints a short-lived ephemeral token server-side, so the real API key never leaves the backend.
• Inngest durable workflows for post-call note distillation and waitlist notifications when a camp goes live.
• Drizzle ORM + PGlite (embedded Postgres) for zero-setup local data, swappable for Neon in production.
• Leaflet + OpenStreetMap for the demand heatmap, with privacy-preserving ~11km aggregation cells.
• Tailwind CSS for an interactive UI: filterable map, trend stat cards, a demand-by-need chart, a camp-request flow, and a redesigned traceability view with a journey stepper and audit timeline.

Challenges we ran into

• Privacy vs. usefulness: Showing NGOs enough to act without ever exposing an individual — solved with coarse geo-cells and minimal data capture (partial phone, region, high-level problem).
• Trustworthy AI: Making every AI decision auditable led us to an append-only design for recommendations and audit events, so nothing the system does is invisible afterward.
• Tooling gotchas: PGlite's WASM assets broke under the bundler until we marked it as a server-external package, and we had to pin the right Node version for the framework. Single-process file-backed DB also meant being careful about migrations vs. a running server.
• Demoability: We built a text-intake fallback that drives the exact same tools and data, so the whole product is demoable even without API credentials.

Accomplishments that we're proud of

• A genuine closed loop: need → recommendation → deployment → notification, end to end.
• AI you can hold accountable — full, human-readable traceability of every step.
• A polished, interactive NGO dashboard (live map, filters, trends, camp workflow) instead of a static report.
• Privacy-first by design, and fully functional with or without live credentials.

What we learned

• In sensitive health contexts, trust is the product — traceability and consent matter as much as the triage itself.
• Aggregation and minimal data capture can deliver real operational value without compromising privacy.
• Durable, append-only workflows make AI behavior debuggable, explainable, and safe to ship.
• Designing for an offline/fallback mode forces cleaner architecture and a far more reliable demo.

What's next for Sakhi: BestFriend

• Real telephony: Integrate with phone/SMS gateways so callers reach Sakhi on any basic handset, no smartphone required.
• More languages & dialects, plus a low-literacy, voice-only experience.
• Proactive follow-ups: Post-camp check-ins and medication/appointment reminders.
• Smarter routing: Predictive demand forecasting and optimal camp placement across multiple NGOs.
• Provider integrations: Two-way coordination with clinics and health workers, and outcome tracking to measure real impact.
• Safeguards & governance: Clinical review of escalation logic, bias monitoring, and consent/audit controls ready for partner deployment.

Built With

• cursor
• grok
• inngest
• vercel
• xai