Automated research → polished report, end to end. Drop in a URL or file. Get back a fully rendered statistical report with narrative, tables, and charts.
| 🟢 Status | Active Development |
| 👤 Owner | Justin Verlin |
| 📅 Last updated | April 9, 2026 |
| 🔗 Repo | — |
| 🧱 Stack | Python 3.10 · Quarto · litellm · OpenAI |
| 📤 Outputs | HTML · PDF · DOCX |
The pipeline has five stages. Each stage is independent and can be swapped out.
Source (URL or file) → Parse → Chunk → LLM Extract (map-reduce) → LLM Report Writer → Quarto .qmd → HTML/PDF/DOCX
| Source (URL or file) → Parse → Chunk → LLM Extract (map-reduce) → LLM Report Writer → Quarto .qmd → HTML/PDF/DOCX |
|---|
| Stage | What happens |
|---|---|
| Input | Detects URL vs. local file, downloads if needed, identifies MIME type |
| Parse | Routes to the right parser — PDF, DOCX, XLSX, CSV, web, or plain text |
| Chunk | Splits text into overlapping context-window-sized pieces; tables kept whole |
| Extract | Each chunk hits the LLM for structured data (entities, stats, claims); results are merged |
| Generate | Second LLM call turns the JSON extraction into a full .qmd with prose + charts |
| Render | Quarto compiles .qmd → HTML / PDF / DOCX |
| Format | Parser | Notes |
|---|---|---|
| 🌐 Web page | trafilatura | Any public URL |
| PyMuPDF + pdfplumber | Text and tables | |
| 📝 Word | python-docx | .docx |
| 📊 Excel | pandas + openpyxl | .xlsx |
| 🗂️ CSV | pandas | |
| 📃 Plain text | built-in |
Add a row each time a new report is generated. Link the
.qmdsource and the rendered output.
| Report | Source | Format | Generated | QMD | Output | Status |
|---|---|---|---|---|---|---|
| Q1 Market Analysis | market_data.xlsx |
HTML | Apr 7, 2026 | 📄 Link | 🔗 Link | ✅ Done |
| Industry Trends | bloomberg.com | Apr 5, 2026 | 📄 Link | 🔗 Link | ✅ Done | |
| Competitive Landscape | comp_data.pdf |
DOCX | Mar 31, 2026 | 📄 Link | 🔗 Link | ✅ Done |
| (next report) | 🔲 Pending |
💡 Tip: Turn this table into a Notion database for filtering by format, date, or status.