![\"team](\"/static/9ada3efdf9ad2afc697d438bd8857754/b15c0/team_pic_with_mentor.jpg\" "\\"team")
\n \n \nTeam with mentor, Vinay Sudhakaran on left
\nSamanvay Foundation | AI Cohort 2 | June 2026
\nAvni is an open-source, mobile-first field data collection platform. It works offline, syncs when there is a signal — even days later — and is sector-agnostic: health, education, sanitation, livelihoods. It is a recognised Digital Public Good (DPG), stewarded by Samanvay Foundation since 2016.
\nToday, Avni powers:
\nBut there has always been a bottleneck: getting an NGO from programme plan to working app takes 2 to 3 weeks of expert engineering time. That cost, paid by every new partner, is what we set out to reduce.
\nMeet Priya, an implementation engineer at Samanvay. When an NGO comes to Avni, she translates their programme requirements into a working app — by hand.
\nHer typical process:
\nThat last part is the most repetitive. A single spec line like \"Last Menstrual Period — cannot be a future date\" becomes:
\nif (lmp > today)\n return \"LMP can't be in the future\";\nMultiply that by ~300 rules per app. Priya writes each one from scratch.
\nThe question we brought to the AI cohort was simple: What if the specification just got converted to a working app automatically — and Priya reviewed instead of typing everything from scratch?
\nSamanvay was selected for Tech4Dev's AI Cohort 2, which began in March 2026. We entered with a clear bottleneck identified and an ambitious goal: automate the journey from programme specification to deployable Avni app using AI.
\nThe path was not straight.
\nWe started with Dify for orchestration. It gave us speed to prototype, but not the flexibility we needed for the nuanced, multi-step generation process Avni apps require. The cohort's midpoint check-in — a structured pause to reflect on progress — became a turning point. We stepped back, acknowledged what was not working, and explored two parallel approaches:
\nWe chose LangGraph. In a month of focused work after that pivot, we had a working pipeline going from specification to app and had started an internal pilot.
\nThe key technical learnings from the journey:
\nAnd one soft skill that turned out to matter just as much:
\nThe final presentation was in front of funders and ecosystem partners — many of them non-technical. The ability to translate the depth of the problem and the complexity of a solution into a story a non-technical listener can feel — and care about — is something I learnt and improvised based on inputs and feedback.
\nFor Ekam's maternal health programme — 8 visit types, 4 programmes, 32 forms, ~300 rules — the only thing that changed was time:
\n| \n | Without Avni AI | \nWith Avni AI | \n
|---|---|---|
| Time to first version | \n2–3 weeks | \n~1 day | \n
The ~300 rules Priya once hand-wrote over weeks are drafted on day one. She reviews and refines instead of typing from scratch. Overall effort reduction: 30–50% per implementation.
\nThe cohort also surfaced another dimension of Avni's AI story: inference at the point of care.
\nAvni now embeds an oral cancer screening model (developed by TANUH at IISc) directly into the field app. A community health worker photographs a patient's oral cavity. The AI — running entirely on-device, offline — returns a verdict: Suspicious or Non-Suspicious. Suspicious cases get referred immediately, without waiting for network connectivity or a specialist visit.
\nIndia sees 140,000+ new oral cancer cases every year. Most are caught too late. Running AI at the edge, inside the same app field workers already carry, changes that equation without requiring new hardware or internet access.
\nThe cohort was as much about community as it was about technology. The sessions and peer exchanges were genuinely inspiring.
\nWatching Pinky's Promise tackle women's health challenges through AI, or seeing how Dalberg is enabling BRAC to deliver skilling to unemployed youth in Bangladesh — these were not just interesting use cases. They were reminders of why this work matters and how different the entry points into social-sector AI can be.
\nTechnical sessions on guardrails and evals gave the cohort vocabulary and practice to build reliably and responsibly.
\nEach NGO team were assigned external mentors for periodic guidance.
\n\n \n \nTeam with mentor, Vinay Sudhakaran on left
\nThe impact of faster implementation is not just about Samanvay's efficiency. It compounds.
\nAnd because Avni is open-source infrastructure — a shared digital public good — every improvement to Avni AI benefits every NGO on the platform, for free. Built once, compounded everywhere.
\nIf you run an NGO and are looking for an offline-capable, open-source field data platform — or if you are building in the social sector and want to explore what Avni AI can do — reach out.
\nAvni: avniproject.org · github.com/avniproject
\nSamanvay Foundation: samanvayfoundation.org
\nSamanvay Foundation builds open-source digital public goods for the social sector. Avni is used by 70+ organisations to track 11M+ beneficiaries across 100M+ field visits.
","frontmatter":{"date":"June 23, 2026","title":"From Weeks to a Day: How AI Is Changing the Way NGO Field Apps Are Built","author":"Maha Lakshme S","description":null,"tags":["AI","AI cohort"],"image":null},"fields":{"slug":"/blog/2026-06-23-ai-cohort-2-final/"}}},"pageContext":{"id":"ad4dc71c-8c63-53be-9eb7-ddc4f8873115"}},"staticQueryHashes":["1175525803","4080856488"]}