Case study · Knowledge platform & AI practice companion
AI Insights: the archive became a working system.
243+ articles stopped being a chronological publishing history and became a searchable, conversational practice companion — with source dates, voice control, playbooks, topic exploration and a structured business-triage mode.
The problem
A large body of useful writing becomes harder to use as it grows. Readers can search titles, but the real value is buried across articles, examples, architectures and trade-offs written at different points in time.
Make accumulated thinking queryable.
The goal was not another chatbot in front of a blog. It was to turn the writing into a practice layer: let readers ask natural-language questions, inspect the sources behind an answer, navigate by topic or role, and move from reading into action.
The platform also had to preserve a basic editorial truth: an old article can still be useful, but the reader should be able to see that it is old.
What changed
- Archive
- 243+ Medium articles imported into a browsable knowledge base.
- Retrieval
- Semantic Q&A across the corpus with source citations and publication dates.
- Interaction
- Text, voice, playbooks, bookmarks, topic graph, quizzes and related reading.
- Action
- A 5-Fuse business-triage mode turns a conversation into a structured diagnosis and 90-day action plan.
What is live
The public product combines discovery, conversation and reusable practice tools rather than treating each as a separate destination.
AI Practice Companion
Ask questions across the full article archive, continue the conversation, and inspect source chips with article titles and publication dates.
Six starter prompts · multi-turn chat · resettable conversations.
Voice as a controller
OpenAI Realtime API voice input supports freeform dictation, explicit submit commands and hands-free shortcuts. Voice drives the same React handlers as the on-screen controls.
One interaction model, two input modes.
Explore rather than scroll
Topic graph, list view, role filters, search, playbooks, bookmarks, related articles and tag drill-down make the corpus navigable in several ways.
Discovery is structural, not just chronological.
Business in trouble?
The 5-Fuse Business Triage runs a short intake, scores RED / AMBER / GREEN fuses, identifies The One Thing, and produces a 90-Day Action Plan that can be exported or privately shared.
A reading platform that can shift into a decision workflow.
The reliability choices
- Freshness is visible.Every source chip includes a publication date so the reader can judge whether an answer is grounded in current or older material.
- Shared control path.Voice commands call the same interface handlers as visible buttons, reducing the chance that hands-free navigation becomes a second, inconsistent product.
- Share the distilled insight, not raw output.Share cards use a separate AI-distilled summary so exported content is concise and self-contained.
- Private operational intelligence stays private.The Writing Relevance Map is admin-only, separating public discovery from internal portfolio analysis.
Under the hood
Selected live mechanisms from the current build.
- OpenAI Realtime API with WebRTC VAD for voice interaction
- DeepSeek used for share-summary generation and weekly digest curation
- Listmonk double opt-in newsletter workflow with anti-spam protections
- HN Algolia + arXiv trend signals for the admin Writing Relevance Map
- Mobile-first layouts, graph/list fallback and responsive filters
Why it belongs in the case-study set
AI Insights is a useful reference build because it joins product design, retrieval, voice, content operations and advisory logic in one system — and it is used in public, where weak navigation, stale sources and vague answers cannot hide behind a demo.
The broader lesson: a knowledge archive becomes more valuable when the system exposes provenance, supports several paths into the same underlying corpus, and turns selected conversations into structured action rather than stopping at an answer.