Turn a corpus of Markdown blog posts into an autonomous marketing knowledge graph that plans, ranks, recommends, and generates campaigns.
This is not a social media post generator. This is a reasoning engine that understands relationships between articles, products, repositories, audiences, and marketing channels — all running locally with no cloud dependency for core operation.
Markdown Corpus → Ingestion → Semantic Analysis → Knowledge Graph
│
Campaign Planner ◄───────┼─── Ranking
│ │
Agent Interface ◄── Recommendation Engine
All structured data lives in a single SQLite database. Vectors live in ChromaDB. The graph is stored as adjacency lists in SQLite with JSON snapshot exports. Every LLM inference is persisted with full reasoning trace.
| Phase | Component | What It Does |
|---|---|---|
| 1 | Ingestion | Parse frontmatter, extract headings/images/links/code, store normalized |
| 2 | Semantic Analysis | Score 27 marketing dimensions per article (LLM) |
| 3 | Topic Extraction | Build normalized taxonomy across 13 categories |
| 4 | Entity Recognition | Extract people, companies, repos, products, technologies |
| 5 | Knowledge Graph | 17 relationship types, typed edges with weights |
| 6 | Duplicate Detection | Exact, near-duplicate, outdated content detection |
| 7 | Marketing Ranking | Weighted composite scores across 12 dimensions |
| 8 | Audience Mapping | 12 audience personas with relevance scoring |
| 9 | Platform Recommendation | 12 platforms scored per article (LinkedIn, X, Dev.to, etc.) |
| 10 | Campaign Planner | Multi-step campaign generation with schedules |
| 11 | Content Repurposing | 10 target formats (thread, newsletter, talk, workshop, etc.) |
| 12 | Marketing Memory | Append-only reasoning history, never regenerates identical decisions |
| 13 | Analytics Schema | Ready for metrics import (views, clicks, conversions, sales) |
| 14 | Recommendation Engine | 11 query types: best today, hidden gems, underutilized, needs update, etc. |
| 15 | Agent Interface | 10 structured tools for autonomous agents + optional MCP server |
| 16 | Autonomous Loop | Nightly: ingest → graph → score → campaign → report |
# Install
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
# Phase 1: Ingest blog posts
python3 -m src.cli ingest
# Phase 5: Build knowledge graph
python3 -m src.cli graph
# Phase 7: Compute marketing rankings
python3 -m src.cli rank
# Phase 14: Query recommendations
python3 -m src.cli recommend today
python3 -m src.cli recommend update
python3 -m src.cli recommend gems
# Full pipeline (requires Ollama)
python3 -m src.cli pipelineLLM-powered phases (semantic analysis, topic extraction, entity recognition, audience mapping, platform recommendations, campaign generation, content repurposing) require an Ollama instance or OpenAI-compatible endpoint. Configure in configs/amis.yaml.
| Component | Choice | Rationale |
|---|---|---|
| Language | Python 3.11+ | Broad ecosystem, async support |
| Structured storage | SQLite | Single-file, zero-config, portable |
| Semantic search | ChromaDB | Local vector store, HNSW indexing |
| Graph model | Adjacency list in SQLite | Separate from documents, exportable |
| LLM reasoning | Ollama / OpenAI | Local-first, swappable backend |
| Markdown parsing | markdown-it-py + python-frontmatter | Full AST, frontmatter support |
| Embeddings | Sentence Transformers | Local, offline-capable |
- Local-first — no cloud dependency for core operation
- Markdown is source of truth — all intelligence derives from authored content
- LLM only where reasoning is required — parse deterministically, reason selectively
- Every inference stored — append-only reasoning trace with model, prompt, and confidence
- Idempotent ingestion — same input always produces same output
- No UI assumptions — pure structured API for autonomous agents
amis ingest # Parse and store all markdown files
amis analyze # Run semantic analysis (LLM)
amis topics # Extract topics (LLM)
amis entities # Extract entities (LLM)
amis graph # Build knowledge graph
amis graph-export # Export graph as JSON
amis duplicates # Find duplicate content
amis rank # Compute marketing rankings
amis rankings # Show ranked articles
amis audiences # Map audience personas (LLM)
amis platforms # Platform recommendations (LLM)
amis campaign # Generate campaign (LLM)
amis repurpose # Content repurposing (LLM)
amis pipeline # End-to-end pipeline (LLM)
amis nightly # Autonomous loop (LLM)
amis recommend <type> # Query recommendations
amis tool <name> # Agent tool interfaceTen structured tools for autonomous agents:
tools.find_best_articles(platform="LinkedIn")
tools.generate_campaign(goal="book_sales", audience="developers")
tools.recommend_platform(article_id=42, top_n=3)
tools.rank_articles(dimension="authority")
tools.find_hidden_gems(min_score=70)
tools.find_duplicate_content()
tools.find_missing_topics()
tools.recommend_book_marketing()
tools.recommend_consulting_content()
tools.generate_monthly_plan()amis/
├── src/ # 2,400 lines across 31 modules
│ ├── cli.py # CLI entry point
│ ├── config.py # YAML config management
│ ├── db/ # SQLite connection + 15-table schema
│ ├── ingestion/ # Phase 1: markdown parsing
│ ├── analysis/ # Phases 2-4: semantic, topics, entities
│ ├── graph/ # Phases 5-6: graph + duplicates
│ ├── marketing/ # Phases 7-11: ranking, audiences, platforms, campaigns, repurposing
│ ├── memory/ # Phase 12: reasoning history
│ ├── recommendations/ # Phases 14-15: engine + agent tools
│ ├── llm/ # Ollama/OpenAI client + prompts
│ └── loop/ # Phase 16: autonomous pipeline
├── content/posts/ # 134 blog posts (canonical corpus)
├── docs/ # 23 implementation documents
├── configs/amis.yaml # System configuration
└── pyproject.toml # Package definition
MIT