The reflections have been put together and shared by Sahaj Parikh from Udhyam (sahaj@udhyam.org)

About the org

• Udhyam – https://udhyam.org/shiksha/
• Program in Focus: Business Projects with adolescents from government schools
  • Phase 1 – 6 weeks – Coming up with an idea and forming a team
  • Phase 2 – 8 weeks – Implementing the idea with seed grant from the government (Rs. 2000 per student)
• Problem Statements:
  • Phase 1 (Coming up with an idea + Forming a team)
    • Idea – Is our idea any good?
    • Submissions – Do our submissions deserve seed money?
  • Phase 2 (Implementing the idea utilizing seed grant)
    • Inputs on how to get started and around basic business concepts
    • Feedback on work – how they are doing – Including acknowledgement (rewards / incentives), course correction and guidance for next steps
      • Personalized and semi-personalised inputs, beyond teacher’s capability
• Use case 1: Support (self help) for teachers and students
  • To help students get started – prototype, market research, first sale
• Use case 2: Feedback on student submis
  • Transcribe audio responses to quiz questions, Evaluate responses using LLM

Top Takeaways

• Focus on creating value first, and then on cost/scale
• Corollary – Focus on the prompt! If the LLM is not behaving as expected, look at the prompt first.
  • A good idea is to categorize queries and send only relevant queries to a knowledge base
  • Learnings from SensAI – for insights shared by Aman (HyperVerge Academy)
  • Prototype and iterate the prompt on OpenAI playground before starting any development effort

Prototyping Done

• LLM to evaluate transcripts of audio recordings – prototyped prompt on ChatGPT playground successfully
• Structured the problem statement for self help use case – will be prototyping it on the playground next

Next steps / Help needed

• Building a solution for an open chat interface, and that too on WhatsApp is trickier as compared to doing it on a custom interface. We’ll need help from glific here, and it will make sense for glific to invest as most of the challenges will be similar for many NGOs
• A best practice for an FAQ type flow (as shared by Aman) is to first categorise the question and then respond according to the type of question. This would require a middleware that makes 2 (or maybe even 3) OpenAI API calls and then responds to the user. Glific can help build a feature for this 

Overall thoughts on the sprint

• I love the concept of getting together with glific, other NGOs (who are in the same boat) and experts! There’s a lot of sharing of ideas, brainstorming, prototyping that happens and I went back with a lot of insights
• Tejas was a great facilitator and anchor – he has been in touch with us for over a month, discussing the use cases, sharing suggestions and following up on actions from our side to ensure that we come to the sprint after having done some work, ready to absorb and implement new ideas. Grateful for his support, and looking forward to a deeper collaboration!
• It was wonderful to have Aman (from Hyper Verge Academy) and Edmund (from Agency Fund) share their knowledge and insights on LLMs. They were super open with their reviews and suggestions, and their inputs will definitely help me go to the next level quickly!