Tokens, embeddings, vector stores, and your first vector database written from an empty file.
Five build days, two live sessions, one ship day, one read day. Monday, September 28 to Sunday, October 4.
Most people arrive able to call an LLM API and unable to explain a single thing happening underneath it. Week 1 closes that gap in the least glamorous way possible: you write the pieces yourself.
By Sunday you will have built a tokenizer and a vector database. Neither is production grade. Both permanently remove the mystery from the tools you will use for the rest of your career.
Nothing this week needs a GPU, a framework, or more than a hundred lines of Python.
| Day | Date | Focus | Type |
|---|---|---|---|
| 1 | Mon, Sep 28 | ‣ | Build |
| 2 | Tue, Sep 29 | ‣ | Build |
| 3 | Wed, Sep 30 | ‣ | Build plus Live Session 1 |
| 4 | Thu, Oct 1 | ‣ | Build |
| 5 | Fri, Oct 2 | ‣ | Build plus Live Session 2 |
| 6 | Sat, Oct 3 | ‣ | Ship Day |
| 7 | Sun, Oct 4 | ‣ | Read Day |
Session 1, Wednesday September 30, 7:30 PM IST, 90 minutes. We write a tokenizer from scratch and run the strawberry test.
Session 2, Friday October 2, 7:30 PM IST, 90 minutes. We build TinyVectorDB, storage, search and persistence.
Tech Support runs every day at 9:00 PM IST on Discord, thirty minutes, drop in as needed.
A cost-tracking CLI, a temperature playground, a working tokenizer, a cosine similarity function written without NumPy, and a file-backed vector store with tests and a README. All of it public on GitHub, all of it committed on the day it was built.
That is Portfolio Build 1 of four.
Day 4 asks you to write cosine similarity in pure Python, no NumPy, no math library. People skip the restriction and import NumPy.
Do not. The restriction is the lesson. Ten minutes of writing the formula by hand is what turns a vector database from magic into arithmetic.