LP–GP Matching System — Vineyard Ventures

Scores Vineyard's LP CRM against a specific GP opportunity and produces a ranked shortlist with outreach recommendations.

---

For reviewers — just open the notebook

All data is pre-cached. No API keys or scraping needed.

pip install -r requirements.txt
jupyter notebook lp_shortlist_analysis.ipynb

Run all cells top to bottom. The notebook loads from data/lp_signals.json and data/raw_lp_data.json — both already present in the repo.

To change the scoring lens (e.g. prioritise geography over sector): edit DEFAULT_LENS in Cell 1 and re-run. Options: deeptech_india (default), sector_first, geography_first, engagement_first, balanced.

---

Files

File	What it is
lp_shortlist_analysis.ipynb	Main deliverable — ranked shortlist, charts, sensitivity analysis, outreach plan
writeup.md	Approach document (submitted as PDF)
data/lp_signals.json	Pre-extracted LP signals (13 LPs, 14 fields each)
data/raw_lp_data.json	Raw Notion data including full call notes
scripts/fetch_notion_data.py	Step 1 — scrapes LP records from Notion via Playwright
scripts/extract_signals.py	Step 2 — AI signal extraction via Cerebras
requirements.txt	Python dependencies
.env.example	Environment variable template

---

Re-running the pipeline from scratch

The two scripts are only needed if you want to re-scrape Notion or re-extract signals. This requires API credentials and is not needed to view results.

cp .env.example .env          # add CEREBRAS_API_KEY
playwright install chromium   # one-time browser install (~150 MB)

python scripts/fetch_notion_data.py   # scrape Notion → data/raw_lp_data.json
python scripts/extract_signals.py     # AI extraction → data/lp_signals.json
jupyter notebook lp_shortlist_analysis.ipynb

extract_signals.py is incremental — it skips LPs already in lp_signals.json, so partial runs are safe.

---

Design decisions

• Playwright over Notion API — the CRM was shared as a public guest link with no API access. Production swap: replace fetch_notion_data.py with Notion API calls; everything downstream stays the same.
• AI for extraction, rules for scoring — LLM handles synonyms and implicit signals in free-text notes; deterministic weighted scoring produces auditable, reproducible rankings.
• Configurable lenses — five weight presets so the same pipeline re-ranks correctly for a different fund type (e.g. geography-first vs sector-first).
• Sensitivity analysis — ranks each LP under all five lenses; an LP that stays top-5 across every lens is a high-confidence pick.