Sprint Projects
This guide outlines our “Sprint Projects”. These are high-intensity, gamified coding sessions. We cover the team structure, the shift to agentic workflows, and the specific challenges you will face in each module.
The Sprint Structure
We use Sprint Projects to turn theory into muscle memory. You work in pairs as Driver and Navigator. You have a strict 60–90 minutes to solve a specific challenge. Post-Module 3, the rules change. You must stop writing code manually. Instead, you act as an “Architect” and use LLMs to generate all syntax. Winners of each sprint earn a spot in the course “Hall of Fame”.
Module 1: The Data Scientist’s Toolkit
“The Tidy Data Escape Room” pits you against a disastrous CSV. It is filled with merged cells and color-coded metadata. Your mission is to clean this dataset into a “Tidy” format within a shared repository. The twist is strict. You must commit and push every single edit. The team with the cleanest data and the most granular git history wins the escape.
Module 2: Visualizing Complexity
“The Ugly Graph Makeover” challenges your aesthetic eye. We give you a misleading, hideous chart. Think 3D pie charts or truncated axes. You must use Python to refactor it into three distinct visualizations that tell an honest story. For example, create views for distribution, relationship, and composition. The class votes on the “Most Insightful” and “Best Aesthetic” results.
Module 3: Agentic Coding
In “Speed-Run: Dashboard from Scratch,” you receive a raw dataset and zero starter code. Your task is to build a fully interactive Streamlit or Shiny app in under an hour. The catch is significant. You are forbidden from writing code manually. You may only prompt the LLM. The first team to have a working local app wins the race.
Module 4: Deep Learning for Text
“The Vibe-Check Classifier” explores semantic space. You start with unlabeled text like Tweets. You must generate embeddings and project them onto a custom Semaxis. Examples include “Engineering vs. Art” or “Happy vs. Sad”. The goal is to identify the “weirdest” outlier. This is the data point the model finds most confusing. The team unveiling the funniest or most counter-intuitive relationship takes the prize.
Module 5: Deep Learning for Images
“Adversarial Art Attack” tests your intuition for CNNs. You are given a pre-trained ImageNet classifier and a set of images. Your task is to modify an image until the model confidently misclassifies a banana as a toaster. You might do this by adding noise via code or digitally drawing on it. The victory goes to the team that tricks the AI with the least visible modification.
Module 6: Deep Learning for Graphs
“The Network Saboteur” asks you to think like an attacker. Given a graph like the Karate Club network, you must identify the minimum number of edges to remove. The goal is to severely degrade the network’s efficiency. You must predict the “load-bearing” edges before calculating centrality metrics. The most efficient destroyer of structure wins.
Module 7: Representations & Structuralism
“The Les Mis Identity Crisis” contrasts two forms of meaning. We provide the raw text of Les Misérables and a network constructed from page-level character co-occurrences. Your task is to generate two rival embeddings for the cast. Create one semantic space using Word2Vec or SBERT on the text. Create a second structural space using Node2Vec on the graph. Compare these spaces via PCA or relative distances to find a conflict. Who is Valjean close to in the story versus the structure? The victory goes to the team that finds the biggest “reputation gap”. We are looking for the character whose narrative vibe most strongly contradicts their network reality.