ResearchRaven
AI-powered research-to-content pipeline that processes knowledge in three stages: Research (flat) → Knowledge Organization (hierarchical) → Content Generation (linear). ResearchRaven automates the journey from source collection to publication-ready content such as articles, newsletters and book chapters. The architecture is grounded in communication theory (Shannon, Saussure, Miller) and the principle that hierarchical knowledge must be linearized for human consumption. The knowledge layer was later rebuilt into two separate concerns — structuring a single document, and linking across documents — after measured evaluation showed that merging both into one graph produced irreversible errors. Knowledge is kept in plain files under version control rather than a database, so it can be rebuilt from source at any time. Built entirely using the AI Development Framework with Claude Code. A dedicated page covers the architecture, the evaluation approach and the vision: /researchraven
Start Date
February 10, 2026
End Date
Ongoing
Duration
6+ months
Tickets Created
292
Tickets Completed
194
AI Sessions
41
- ●✅ 3-stage architecture grounded in communication theory
- ●✅ Competitive research on 10 AI research tools before writing any code
- ●✅ Deep research on linearization theory (82 sources)
- ●✅ Hierarchical knowledge organization across many sources
- ●✅ Configurable writing styles for several output formats
- ●✅ Post-processing pass for natural phrasing
- ●✅ First end-to-end run: research topic → publishable draft
- ●✅ Incremental updates so recurring research does not rebuild everything
- ●✅ Multi-perspective research covering technology, business and research angles
- ●✅ Automated source-quality assessment
- ●✅ Schema validation for every stored knowledge format
- ●✅ Explicit linearization step bridging hierarchy to linear content
- ●✅ Comprehensive test coverage per pipeline component
- ●✅ Acceptance thresholds fixed in decision records before measurement
- ●✅ Blind evaluation with an independent judge after a same-judge result proved misleading
- ●✅ Knowledge layer rebuilt on measurement rather than tuned to pass
- ●✅ Derived state fully rebuildable from source files, byte-identical
- ●✅ Explicit cost governance as the pipeline grew
- ●✅ Validated AI Development Framework for AI pipeline projects
This project follows a multi-repository architecture with specialized repositories for different concerns.
Meta Repository
Private Repository
App (Orchestration)
Private Repository
Research (Stage 1)
Private Repository
Knowledge (Stage 2)
Private Repository
Content (Stage 3)
Private Repository
Knowledge Linking Layer
Private Repository
Source Fetcher
Private Repository
Storage Layer
Private Repository
AI Conversations
Private Repository