Guruthuma — The Teacher's AI Counterpart

The Bottleneck

One teacher, one thousand doubts.

this is the whole thesis

Sri Lanka's best tuition teachers are already at capacity. Every evening, WhatsApp fills with the same handful of doubts — repeated across hundreds of students who won't get a reply until class next week. The teacher's time is the product, and it does not scale. Guruthuma exists to answer the repeatable 80% instantly, in the teacher's voice, so the teacher's actual attention goes to the 20% that genuinely needs it.

What breaks today

Doubts pile up faster than one person can answer
Generic AI chatbots answer in stiff, translated Sinhala students don't trust
Good explanations get typed once, then lost in chat history forever
Teachers can't tell which questions are genuinely new

Two Approaches

Pick your entry point.

A proves it works.
B makes it a business.

Approach A

The Telegram Sandbox

— fast, free, built to learn from

GoalProve natural Sinhala + accuracy with one real teacher

PlatformTelegram Bot API — free, no approval wait

Cost to runLLM tokens only, no messaging fees

Setup timeDays, not weeks

Best forValidating tone, testing one subject, one teacher pilot

Start here. There is no reason to spend on Meta's messaging fees before the Sinhala tone is actually convincing.

Approach B

The WhatsApp Production Suite

— where the students already are

GoalScale across many teachers, many students, as a paid product

PlatformMeta Cloud API (WhatsApp Business)

Cost to runPer-conversation fee on user-initiated 24hr windows

Setup timeWeeks — business verification + Meta review

Best forThe subscription product teachers actually pay CodeShop for

Move here once Approach A proves the answers hold up — this is the version you sell, not the version you test with.

Core Architecture

Four moving parts.

keep it this lean for v1

The Interface

Telegram Bot API → WhatsApp Cloud API

Where the student actually types the question. Same backend serves both once Approach B kicks in — the channel is swappable, the brain underneath isn't.

The Brain

Fast, high-context LLM (Gemini Flash-class / GPT-4o-mini-class)

Needs to be quick and cheap per message, with enough context room to hold the teacher's tone guide and syllabus excerpt for every single reply.

The Knowledge Base

RAG pipeline · vector database (Pinecone / Supabase)

Guruthuma never improvises from the open internet. Every answer is retrieved from that teacher's own uploaded notes, past papers, and marking schemes first.

The Backend Controller

Node.js / Python (FastAPI + LangChain)

The quiet middleman — connects messaging app, vector store, and LLM, and holds the session memory and escalation logic described below.

The Real Hard Part

Sounding like Sir, not Google Translate.

if this fails, nothing
else matters

Default AI Sinhala

Stiff, literary "Likhita" register

Invented technical terms, wrong exam vocabulary

Reads like it was translated, not taught

Flat tone — no encouragement, no authority

Guruthuma's Register

Natural spoken/exam Sinhala, the way class actually sounds

Locked technical glossary per subject (e.g. A/L Physics)

Trained on the teacher's own past WhatsApp replies

Encouraging, direct, quietly authoritative — like Sir

Tactic 01

Few-shot locking

10–20 real student questions paired with the teacher's actual replies, embedded directly in the system prompt.

Tactic 02

Glossary enforcement

A strict per-subject term list so the model uses correct exam vocabulary instead of a literal translation.

Tactic 03

Tone instructions

Explicit direction toward an encouraging, direct, mildly authoritative voice — a tutor's tone, not a search engine's.

Action Plan

From sandbox to subscription.

one teacher, one
subject, first

Week 1–2

Foundation & teacher grounding

Pick one pilot teacher, one subject. Stand up the Telegram bot, a basic upload portal, and chunk their existing notes and past papers into the vector database.

Deliverable → working Telegram bot answering from real notes

Week 3–4

Sinhala tone engineering

Collect 15–20 of the teacher's real WhatsApp replies, build the few-shot prompt and subject glossary, and run it past the teacher directly for correction.

Deliverable → tone-locked prompt signed off by the teacher

Week 5–6

Pilot with real students

Roll out to one class. Add Redis session memory for 24-hour follow-up context, and the escalation protocol for anything outside the notes.

Deliverable → live pilot, teacher reviewing a flagged-question dashboard

Week 7–8

Refine & add semantic caching

Tune accuracy from real pilot data. Add semantic caching so repeated questions are answered instantly without a fresh LLM call.

Deliverable → response cost and latency both down, accuracy holding

Month 3

Move to Approach B — WhatsApp

Apply for Meta Cloud API access, complete business verification, and port the proven backend from Telegram to WhatsApp.

Deliverable → production-ready WhatsApp deployment

Month 4+

Package & sell across teachers

Turn the pilot into a repeatable onboarding flow — new teacher, new subject, new glossary — and open subscriptions beyond the first case study.

Deliverable → second and third teachers onboarded, pricing validated

Business Model

Who pays for what.

Per-teacher subscription

Monthly platform fee

Covers hosting, LLM usage, and the vector database for that teacher's subject library — billed regardless of student volume.

Messaging pass-through

WhatsApp conversation fees

Meta's per-conversation charge on Approach B gets factored directly into the teacher's subscription tier, not absorbed.

Onboarding & setup

One-time build fee

Ingesting notes, building the glossary, and tone-locking the prompt is teacher-specific work — priced as a setup case study, per CodeShop's usual engagement model.