AIPathshala

Problem#

Teachers spend significant time turning textbook content into usable lessons — finding the right material, structuring it pedagogically, and adapting it for their classroom. AIPathshala is an educational SaaS platform built to shorten that gap: ingest a school's actual textbooks, and use retrieval-augmented generation to help teachers generate lessons directly from that material, inside an editor built for the job rather than a generic document tool.

My Contributions#

I worked as part of a small engineering team on AIPathshala, and my scope covered nearly the entire system: the Docling-based textbook ingestion pipeline, the RAG lesson-generation system, the lesson editor, and the Discord-inspired school administration system — with validation being the one piece outside my direct ownership.

Architecture#

At a high level, the platform ingests textbooks via Docling, indexes the resulting content for retrieval, and exposes a RAG-driven lesson-generation flow to teachers through a highly customizable lesson editor — comparable in capability to a full document editor rather than a simple text box. Alongside content generation, the platform includes a school administration layer modeled on Discord's role and permission system, letting schools define who can do what across the platform.

Technical Decisions#

Docling was chosen for textbook ingestion, and retrieval-augmented generation for lesson creation — grounding generated lessons in the school's actual textbook content rather than generating from a model's general knowledge alone. The administration system's role/permission model was deliberately modeled on Discord's, rather than a conventional flat admin/teacher/student scheme, to support the more varied structures real schools actually have.

Engineering Challenges#

The RAG lesson-generation pipeline was the harder of the platform's two major subsystems (the other being the Discord-style admin/permissions system). Getting the generation agent itself to a reliable standard took real iteration — it started out fairly mediocre and became a genuinely reliable system only over time, through repeated tuning. The lesson editor's formatting also went through extensive iteration and feedback cycles before it reached the level of polish a real document-editor-grade tool needs.

Lessons Learned#

Formatting and editor polish took far more iteration than expected — building something that feels like a real document editor, rather than a rough text box, is genuinely tedious work with many rounds of feedback. On the generation side, the clearest lesson was that a RAG-driven agent's first working version is rarely its final quality bar: what started as a fairly mediocre generator became reliable only through sustained iteration, not a single design decision.

Tech Stack#

Docling for textbook ingestion, and a retrieval-augmented generation pipeline for lesson generation. Further implementation detail is not shared here, consistent with this project's proprietary status.

Related Content#

TextStruct — a personal project originally built as an experiment for AIPathshala, evaluating whether pedagogical role classification could improve retrieval quality over conventional chunking. That experiment's conclusions directly informed this platform's ingestion architecture.

SalesTaxGST — another AILACDS professional project, also built on Docling-based ingestion and retrieval-augmented generation, for a different domain (tax consulting rather than education).