Learnings from OpenAI's event on AI for Social Impact

Last week, I attended OpenAI's event on AI for Social Impact in Bangalore.

Engineers from OpenAI gave a walkthrough of their new Agents SDK and shared what to expect in the coming months. This was followed by lightning talks from non-profits like Rocket Learning, Noora Health, Udhyam Learning Foundation and others, showcasing how they are already putting AI to use on the ground.

I spoke about SensAI, an AI-powered LMS we are building to help educators provide real-time, personalized support to learners while reducing their own workload. I also shared some of the learnings from our journey so far.

OpenAI also launched OpenAI Academy in India, a learning platform for individuals and teams to apply AI in their work (from basics to advanced).

Some reflections from the day:

๐Ÿ. ๐‚๐จ๐ฉ๐ข๐ฅ๐จ๐ญ๐ฌ ๐ž๐ฏ๐ž๐ซ๐ฒ๐ฐ๐ก๐ž๐ซ๐ž In high-trust rolesโ€”teachers, healthcare workers, anganwadi staffโ€”people rely on trust, accountability, reassurance, and someone who will tell it like it is. Replacing the human layer rarely works. But giving these professionals tools to be more effective? Thatโ€™s where AI can quietly but powerfully shift outcomes.

๐Ÿ. ๐‘๐ข๐ฌ๐ž ๐จ๐Ÿ ๐ฏ๐จ๐ข๐œ๐ž ๐ข๐ง๐ญ๐ž๐ซ๐Ÿ๐š๐œ๐ž๐ฌ Speaking is faster, more comfortable, and more accessible. In many contexts, it's the most natural way to interact. Thereโ€™s a clear shift underwayโ€”from text-only chat systems to voice-first interfaces.

๐Ÿ‘. ๐„๐ฏ๐š๐ฅ๐ฌ ๐š๐ฌ ๐ญ๐ก๐ž ๐Ÿ๐จ๐ฎ๐ง๐๐š๐ญ๐ข๐จ๐ง Irrespective of whatever advances are made in LLMs, having a robust evals system is crucial for any AI product to succeed. It was inspiring to see many non-profits adopting the evals-first approach tailored to their own domainโ€”sometimes with AIโ€™s help, but always grounded in the real-world needs of their usersโ€”something which was missing an year back.

๐Ÿ’. ๐Œ๐ฎ๐ฅ๐ญ๐ข๐ฅ๐ข๐ง๐ ๐ฎ๐š๐ฅ ๐ข๐ฌ ๐ง๐จ๐ง-๐ง๐ž๐ ๐จ๐ญ๐ข๐š๐›๐ฅ๐ž English is not enough. For AI to matter to most of India, it needs to work well in our native languages. That gap is still largeโ€”and critical to close. Frontier models continue to struggle with low-resource languages, which leads to a real dilemma for many teams: should they wait for the next model update from OpenAI, or start building their own models to address this gap now?

๐Ÿ“. ๐‚๐จ๐ฌ๐ญ ๐ฌ๐ก๐š๐ฉ๐ž๐ฌ ๐ซ๐ž๐š๐œ๐ก Even with prices dropping 50x in the last two years, AI at the last mile in India remains expensive. For Indiaโ€™s scale, OpenAI was urged to explore a more affordable pricing model specific to this context. Evals can help here tooโ€”by deciding when a much smaller, cheaper model is good enough instead of defaulting to the largest one.

After nearly 7 years working at the intersection of AI and social impact, this is the most forward movement Iโ€™ve seen. Now itโ€™s about building with care, clarity, and deep context.

Here are my slides.