Future Work

Things deliberately not in v1, captured so they don’t get lost.

PDF OCR (image-only decks)

v1 behavior: assumes the input PDF has a text layer. Image-only decks (scanned, exported-as-images) won’t yield names/titles via straight extraction.

Status: already solved in memopop-orchestrator. When the need arises, port the OCR pipeline from there rather than reinventing.

Where to look: memopop-orchestrator repo (likely Python, possibly pytesseract + pdfplumber + pdf2image, in a .venv). Confirm the exact stack at port time.

Integration shape (when ported): add a scripts/pdf-ocr.{ts,py} step that runs before the discovery phase, producing a text-layered PDF or a markdown transcript that the rest of the pipeline consumes. Keep the OCR step optional — most VC decks have a text layer.

Brandfetch (SVG logos + brand kits) — DONE in v2 (2026-05-10)

v2 behavior: Brandfetch is tier 4 of the logo cascade in SKILL.md. scripts/brandfetch.ts reads BRANDFETCH_API_KEY from ~/.secrets and exposes three modes:

./brandfetch.ts stripe.com                              # full JSON response
./brandfetch.ts stripe.com --best-svg                   # one-line URL or exit 3
./brandfetch.ts stripe.com --best-raster                # one-line URL or exit 3
./brandfetch.ts stripe.com --cache-dir <path>           # cache by domain

pickBestSvg prioritizes type=logo + theme=light with transparent background. Free tier: 1k req/mo.

Background-strip pipeline — DONE in v2 (2026-05-10)

v2 behavior: scripts/bg-strip.sh runs ImageMagick flood-fill from each corner (sampling the corner color, so it works on white and arbitrary brand-color backgrounds). Falls back to rembg (U²-Net) when flood-fill removes <5% of pixels — caller is expected to install rembg via pipx install rembg[cli] for the tier-2 fallback.

Outputs JSON like:

{"method":"imagemagick-floodfill","bg_color":"#ffffff","pct_transparent":94.4,"output":"./logo.png"}

The CP3 portfolio fetch loop should call this automatically after every raster download, then write logo_bg_stripped, logo_bg_strip_method, and logo_bg_color_detected into the company’s frontmatter.

Inline SVG hunt — DONE in v2 (2026-05-10)

v2 behavior: scripts/logo-hunt.sh runs tiers 1–3 of the cascade locally (no paid APIs):

1. Inline <svg> from the homepage’s nav/header
2. Common SVG paths + any .svg URLs scraped from the homepage HTML
3. Brand / press / media pages

Returns JSON like {"tier":"site-svg-path","url":"https://..."} or exits 2 if all three local tiers miss — caller then escalates to brandfetch.ts (tier 4), Tavily site-search (tier 5), or Google CSE (tier 6, still future).

Google CSE fileType=svg (last-resort SVG search)

v1 behavior: not implemented.

Future: add scripts/google-cse.ts reading GOOGLE_CSE_KEY + GOOGLE_CSE_CX from ~/.secrets. Use fileType=svg parameter. Free quota: 100 queries/day. Only worth the wiring when Brandfetch + Tavily site-search both miss frequently.

Crunchbase API (vs. scraping)

v1 behavior: Tavily search → Firecrawl on the public Crunchbase page. Limited data (Crunchbase paywalls).

Future: if the team subscribes to Crunchbase API, add scripts/crunchbase.ts. Until then, public-page scraping is fine.

Site adapters (mpstaton-site, fullstack-vc, hypernova-site, memopop-ai)

v1 behavior: the skill writes only to <cwd>/data/firms/{firm-slug}/. Wiring into a specific site’s content collection is manual.

Future: per-site adapter scripts that read data/firms/{firm-slug}/ and write to the site’s content collection format. One adapter per site, lives in the site’s repo, not in this skill (because each site’s schema is different).

Confidence-driven re-fetch

v1 behavior: files marked confidence: low or flagged need manual review.

Future: a --retry-flagged mode that re-runs only the flagged entities with a different cascade order (e.g., escalate to Firecrawl extraction with a stricter schema, or try alternate query phrasings in Tavily).

Preview UI

v1 behavior: the user reviews .md files in their editor.

Future: a small Astro page (in design-system-viewer style) that renders data/firms/{firm-slug}/ as a browsable preview — see all team members with headshots, all portfolio companies with logos, flag-status badges. Helps the human review pass go faster.

Source-of-truth de-duplication

v1 behavior: each firm has its own copy of every entity. If Alfred Lin appears under both Sequoia and another firm we work with, two separate team/alfred-lin.md files exist.

Future: optional ~/.claude/skills/crawl-fetch-ingest/people/{global-slug}.md index that firm-level files reference — so updating LinkedIn URL once propagates. Only worth building when the user has 3+ firms and the duplication starts costing time.