EveryDay Learning AI Study App

Problem

Saved articles only become reusable study material when the product can turn them into structured questions, plausible answer choices and well-timed review prompts inside the learning flow.

Constraints

• Question generation had to transform stored articles into app-ready quiz data rather than free-form model output.
• Distractors needed to stay distinct from the correct answer while remaining grounded in the article context.
• Review scheduling had to respect per-user, per-question history and trigger only first-time or seven-day-repeat sessions.
• The work was delivered inside an enterprise product, so no public repository or internal usage metric is exposed here.

My contribution

• Built the article chunking and question-generation pipeline with Hugging Face Transformers and T5, then normalised model output into the structured records consumed by the Android quiz experience.
• Implemented context-aware distractor generation by jointly encoding article content, question and correct answer, plus token cleanup, delimiter parsing, blank handling and case-insensitive deduplication.
• Connected the workflow through a local Flask API and an AWS Lambda path that read pending articles, wrote generated quiz data back to MySQL and closed the article-to-quiz loop.
• Implemented seven-day spaced-review scheduling and learning-progress writes, including SQL aggregation over per-user, per-question history and weekly review summaries for the mobile product.

Architecture / methodology

• Split saved articles into generation-ready segments, run T5-based sequence generation, and reshape the output into question-plus-correct-answer records.
• Use a specialised distractor-generation T5 model with context-aware encoding so multiple-choice options stay tied to the source article rather than generic filler.
• Persist generated questions, review eligibility and learning activity through MySQL-backed services consumed directly by Android quiz and review flows.

Evaluation

• Use the app-facing pipeline itself as the integration boundary: article ingestion, question generation, choice construction, review scheduling and learning-record writes must complete as one chain.
• Check generated choices for token noise, separator issues, blank outputs and case-only duplicates before they are accepted.
• Avoid claiming public quality, accuracy or engagement metrics because no public benchmark or repository evidence is available.

Results

• The delivered workflow connected article ingestion, question generation, multiple-choice construction, seven-day review scheduling and study-record writes as one product-facing learning loop.
• This entry intentionally avoids unsupported usage or performance claims because the project shipped inside an enterprise product without public artefacts.

Evidence and links

• Zhifeng Li worked on the Everyday Learning AI study app as an AI/NLP Development Engineer between November 2023 and April 2024.

  Owner-supplied internship and project summary

• The work covered question generation, distractor generation, seven-day review scheduling, MySQL writes and Android quiz consumption for the enterprise product.

  Owner-supplied internship and project summary