Tianci's Toolbox
AI LAB

AI Apps

Content agents, enterprise knowledge search, and news automation projects I worked on, with details of backend implementation, failure handling, retrieval evaluation, and CMS publishing integration.

Internal use only

Python · FastAPI · OpenAI SDK · PostgreSQL · Langfuse

Financial Content Agent & Publishing Platform

Independently developed the Python / FastAPI backend and integrated search, the database, and the CMS. Reduced the time from starting an article to final publication from roughly 3–5 days of manual production to about one hour.

My implementation and results

  • Separated research, outlining, writing, review, revision, and publishing into steps that exchange structured results, with automatic revision and manual review support.
  • Wrapped search and CMS tool calls with timeout, retry, and fallback handling. Stored task state, sources, and revision history, with task retries and idempotent publishing.
  • Used Langfuse to record prompt versions, token usage, latency, tool calls, and errors, and maintained evaluation and regression checks based on real business cases.
  • Generated titles, summaries, FAQs, internal-link suggestions, and Japanese articles for financial / FX content, then sent reviewed content into the CMS draft and publishing workflow.
  • About 95% of tasks complete the workflow and pass review without manual intervention. A manual review path remains available for content that needs human judgment.

Content production workflow

  1. 1Research: search and source collection
  2. 2Outline: structure and SEO requirements
  3. 3Write: draft from source material
  4. 4Review: facts, wording, and publishing requirements
  5. 5Revise: address review feedback
  6. 6Publish: write to the CMS and publish
Internal use only

Python · FastAPI · LangChain · Milvus · PostgreSQL · Redis

Enterprise Knowledge Base & AI Assistant

Developed document processing, retrieval, and Q&A services that answer from internal documents and return sources, focusing on inconsistent retrieval results and index maintenance when documents change.

My implementation and results

  • Implemented document parsing, cleaning, chunking, and embedding; built Milvus indexes and stored sources, categories, and update timestamps.
  • Combined query rewriting, metadata filters, BM25 and vector search, then reranked candidates to address inconsistent vector-only retrieval results.
  • Generated answers from retrieved content with source references. Evaluated Recall@K, answer correctness, and citation accuracy on real business questions, using failures to adjust chunk size, retrieval count, and search strategy.
  • Used document IDs and content hashes to detect changes and update indexes incrementally, reprocessing only affected content.
Internal use only

n8n · GPT · CMS Automation

Automated News Workflow

Built an n8n workflow for news discovery, content processing, localization, and publishing, reducing copying, switching between systems, and repeated publishing steps.

My implementation and results

  • Searched for current topics, filtered duplicate story ideas, and parsed source articles into structured content.
  • Connected AI rewriting, localization, and cover generation so source material could feed publishing workflows in multiple languages.
  • Refreshed the sitemap after CMS publishing and sent execution reports through DingTalk.
  • Organized repeated steps into reusable nodes for changing sources, processing steps, and publishing destinations.