Week of 7 Sep 2026¶

What happened¶

The call begins with Speaker 1 discussing domestic moving tasks before turning to review the tuition center proof-of-concept system developed by Miyuru. The team analyzes the operational pain points of Sri Lankan tuition centers and schools, exploring attendance tracking, student management, and existing solutions like EduMix. Speaker 1 emphasizes that rather than building a standard Learning Management System or static UIs, the system must be AI-native and focus on creating business insights and saving time for teachers and owners. They identify teachers and tuition center owners as viable target segments for early adoption, while ruling out student-facing monetization due to unsustainable token costs. The architecture discussion highlights the need for an intent-driven model harness using OpenRouter rather than hardcoded workflows, enabling the system to adapt dynamically to non-technical users. Speaker 1 stresses conducting direct discovery interviews with tuition teachers to uncover real operational pain points before committing to code. In the final segment, [name withheld] reports on issues with the website and Pali Tripitaka repository due to Anthropic API key requirements and outdated Groq APIs. The team proposes implementing an open-source Model Context Protocol (MCP) server backed by vector search across the Tripitaka to eliminate expensive continuous Claude API token costs.

Key points¶

Decided · The project will not target students as paying customers for the POC because students cannot pay and would cause unsustainable token cash-burn.
Decided · The primary target audiences for the initial POC will be tuition center owners and teachers, excluding students and parents from the first phase.
Said to exist · Miyuru implemented a tuition center prototype dashboard showing teachers, subjects, and classroom student numbers with dynamic chat suggestions from sample data.
Planned · The team plans to formulate key interview questions and interview approximately ten tuition teachers to discover real operational pain points before developing software.

Ideas and open questions¶

Idea · Teachers should be able to update their syllabus progress and classroom status simply by submitting short voice notes after each class.
Idea · Miyuru should redesign the prototype to avoid hardcoded generated UI code, structuring it instead around a dynamic harness that captures user intent and evolves with user interaction.
Decided · The Tripitaka repository on the website is failing to connect because Claude requires an Anthropic API key and the old Groq API configuration is outdated.

All details (4 topics, 3 more points)

Tuition Center Proof of Concept and Operational Analysis: Miyuru presents his prototype dashboard displaying teachers, classes, and student metrics. The team debates how data can be reliably ingested from physical Sri Lankan tuition institutes, comparing manual paper card marking, barcode/QR readers, and existing center apps.
Competitive Evaluation: EduMix and Sri Lankan EdTech Platforms: The team discusses CCG in Gampaha and existing market systems like EduMix, evaluating their subscription pricing and student/parent portals. They agree their product should not compete as a traditional LMS, but rather act as an AI-native operational and insight layer.
AI Harness Architecture and User Experience for Non-Technical Users: Speaker 1 explains that non-technical teachers cannot be expected to write skills or prompt engineering workflows. He advocates for a flexible model harness connected via OpenRouter that infers intent from raw dumped data or short voice notes, evolving dynamically alongside the user.
Dhamma Tripitaka Repository and MCP Integration: [name withheld] explains that the Tripitaka repository cannot run queries properly without an Anthropic API key, since the previous Groq API integration is outdated. Speaker 1 proposes utilizing an open-source Pali Tripitaka MCP and vector search embeddings to avoid expensive proprietary cloud model consumption.
Idea · The tuition platform should initially launch as a responsive web application that can later be packaged as a mobile wrapper rather than building separate native Android and iOS apps.
Idea · OpenRouter should be used as the model routing backend to allow flexible access to high-end, low-end, and tile models according to task complexity.
Idea · To avoid continuous Anthropic API token billing, the team should implement an open-source Pali Tripitaka Model Context Protocol (MCP) server integrated with vector search embeddings.