Jin Jianzuo

A decision-support platform that helps lecturers identify students needing additional support and generate personalised intervention recommendations. Analytics, conversational AI and recommendation generation sit behind a natural language interface, so lecturers ask questions directly instead of reading across multiple reports.

• Conversational layer. Built on Amazon Bedrock AgentCore, with long-term memory carrying context across sessions and modular tool interfaces for retrieving analytics and recommendation outputs.
• Never reading half-written data. Agents risked resolving partially published datasets. I added a pointer-based commit mechanism using lightweight JSON pointers in S3 — pointers update in a strict sequence and the final state pointer acts as the commit signal, so analytics only ever run on complete, consistent data.
• Numbers the model can't invent. Evaluation showed the LLM occasionally fabricating statistics rather than reporting computed values. Moving all numerical aggregation into deterministic Python tools left the model responsible only for presenting verified results in natural language.

• Amazon Bedrock AgentCore
• Python
• AWS S3
• LLM Agents

Client engagement — implementation details limited by confidentiality.

An AI dictionary for children aged four to ten, built after watching my seven-year-old brother struggle with definitions written for adults. A search returns an age-appropriate definition, pronunciation, example sentences, synonyms, and an AI-generated comic strip illustrating the word in context. Saved words feed quizzes and crossword puzzles.

1 / 3

• Free to operate, by design. Each unique combination of word, age group and prompt version is generated once and served from a shared cache thereafter; comics are compressed to WebP to stretch the free storage quota.
• Cutting authentication entirely. Accounts existed solely to store saved vocabulary, so I replaced them with a local-storage abstraction — no personal data collected, simpler for young users, and the storage layer stays abstracted so cloud sync can arrive later without an architectural rewrite.
• Degrading gracefully. Words are validated against a dictionary API before any model call, so invalid input never triggers a paid request. If comic generation fails the text still returns; if text generation fails nothing is cached, so the next request retries.

• LLMs
• Image Generation
• Shared Caching
• WebP
• Local Storage

Currently an MVP — deploying soon.