A collection of hands-on engineering assignments designed to teach students how to think critically when building with generative AI.
GenAI is great at producing plausible first drafts—and terrible at questioning hidden assumptions. Each module here gives students a realistic, slightly underspecified task that looks easy to "one-shot" with an LLM, then evaluates their solution against a benchmark built from real-world edge cases, dirty data, and pipeline constraints.
01-pii-scrubber/(based ongojiplus/piinata)- Task: Use GenAI to build a PII scrubber for research/survey files.
- What GenAI usually misses: Multi-format reading & silent file-skipping, split
lat/loncolumns, column-name substring traps (reason_visit,latrine_type), PII hidden behind question numbers (q12), dirty Unicode / non-Latin scripts, Qualtrics 3-row headers, and false-alarm decoys (villages, schemes, 9-digit order IDs).