Most teens use group chats to swap memes and homework screenshots. A recent HigherEds bootcamp team asked a different question: what if the group chat itself became a private tutor? By the end of a single month they had shipped a Telegram bot that quizzes users on class notes, tracks weak topics and recommends custom study blocks. The project proves that with a clear sprint plan, open-source tools and on-call 1:1 mentorship, ambitious students can move from idea to deployment in record time.

Week-by-Week Build

The sprint board lived on GitHub Projects; each card linked to commits so mentors could review code asynchronously. That workflow alone saved more than ten hours of back-and-forth and mirrors the agile patterns used in professional AI teams.

Deep-Dive on Key Build Elements

1. Content Pipeline

• Markdown parser strips YAML front-matter, cleans LaTeX, and chunk-splits at 300 tokens to maximise retrieval precision.

• Each chunk stored in Pinecone with two metadata keys: subject and difficulty. This allowed the bot to filter biology vs. calculus on demand.

def mk_chunks(md_file, max_tok=300):
    text = open(md_file).read()
    paras = [p for p in text.split("\n\n") if p.strip()]

2. Embedding & Retrieval

• Selected text-embedding-3-small for its 512-token limit and low cost ($0.02 / 1k tokens).

• Index dimension: 1536, similarity metric: cosine, top-k: 5.

• Retrieval latency averaged 42 ms per query on Pinecone’s Starter tier.

3. Dialogue Manager

• LangChain Retrieval-QA Chain with re-ranking: first aggregate top-k chunks, then prepend them to system prompt.

• Prompt template reserved 1 000 tokens for context, 300 for user question, leaving 200 for answer headroom to stay within GPT-3.5-Turbo limits.

• Added a Confidence Survey (“Was this helpful? 1–5”). Score ≤ 3 triggers follow-up questions tagged in SQLite.

4. Telegram Integration

• One Flask route /webhook receives updates.

• Telegram secret token stored in GCP Secret Manager; loaded via environment variable at runtime.

• Implemented exponential back-off to handle 429 rate-limit errors from Telegram.

5. Adaptive Spaced Repetition

• Each Q&A round logs user_id, chunk_id, confidence, timestamp.

• Every midnight, a cron job runs an Ebbinghaus decay function to schedule the next quiz date.

• Low-confidence chunks surface sooner; mastered chunks push out to 7-day, 14-day, 30-day intervals.

What Made the Bot Smart

1. Context windows that fit teen brains
   Instead of dumping full textbook chapters, the parser chunked notes into 300-token passages keyed by headings. Retrieval then surfaced the smallest chunk that answered a query, keeping responses under thirty seconds of reading.

2. Adaptive questioning
   After every answer the bot tagged user confidence on a five-point scale. Scores below three fed directly into the next quiz cycle, imitating spaced repetition. A pilot test with nine students showed a 28 percent jump in recall on a unit quiz compared to peers using static flashcards.

3. Data privacy
   All embeddings stored only anonymised note IDs, no raw sentences. Parents appreciated that design choice, and it satisfied school policy on sharing class material.

Why the Timeline Worked

• Focused scope: No grades, no gamification, just Q&A plus tracking.

• Mentor guardrails: When the retrieval pipeline looped, a tutor spotted the missing await call in minutes.

• Early users: Recruited ten classmates on day ten, catching confusing prompts long before launch.

HigherEds bootcamps bake these principles into every project. Weekly checkpoints stop scope creep, and one-to-one tutoring fills theory gaps the moment they appear.

Lessons for Future Builders

• Start with high-quality notes; garbage in still means garbage out.

• Retrieval-augmented generation beats giant context windows for speed and cost.

• A simple confidence survey unlocks powerful personalised practice.

• Privacy earns trust; strip PII before you embed anything.

• Demo early and often; classmates are ruthless but invaluable testers.

Where to Go Next

The team is now exploring a progress dashboard that recommends session plans to parents and tutors. They are also testing a “voice chat” mode using Whisper for students who prefer speaking to typing.

If your teen spends evenings scrolling for study help, imagine them improving their own chatbot instead. Consider enrolling them in the next HigherEds AI Bootcamp or booking a block of tutoring hours to turn daily questions into project fuel.