solo-developer

Solo developer creating agent-powered tutoring

PythonAzure AI Foundry (GPT-4.1GPT-4.1 Nanotext-embedding-3-large)Semantic KernelAzure Cloud Services (Cosmos DBAzure AI SearchAzure Container Apps)ChainlitFastMCPDocker
Stack tools11
AddedMar 2026
StatusPublished

Found 2 strong examples: ATLAS (91.1% exam score vs 77.8% baseline) by Master's student Dany; Docker AI Tutor prototype validating embedded low-latency tutoring. Limited exact solo matches; many team/academic/team projects.

solo-developer

Why they built it

General-purpose LLMs like ChatGPT lack student background awareness, course material grounding, progress tracking, and Socratic guidance, limiting their tutoring effectiveness compared to human educators.

What worked

Knowledge Tracing for KC management, chat summarisation and checkpoints for context/memory, RAG for grounded responses, SAILED evaluation framework, Semantic Kernel's flexibility for live prompts.

What broke or was painful

LLMs struggle with long inputs (workaround: KC decomposition, summarisation); preference inference inconsistent; no emotional cues/personality awareness; text-only limitation; simulated students not fully representative.

The result

Found 2 strong examples: ATLAS (91.1% exam score vs 77.8% baseline) by Master's student Dany; Docker AI Tutor prototype validating embedded low-latency tutoring. Limited exact solo matches; many team/academic/team projects.

References