LearnStart, rebuilt — from a lost Heroku site to an Astro Knots front door

Why Care?

LearnStart — the globally-recognized Pre-Seed to Seed arm of Learn Capital, managed independently — lost its website when the original Heroku app went away. A firm that has backed over 60 companies transforming the way the world learns since 2015 shouldn't be a broken link.

So we rebuilt it, and we rebuilt it to be un-loseable: an Astro Knots site that deploys from its own repo to Vercel. No single platform's sunset can take it down again — the whole thing is a folder of Markdown, data, and .astro files we own.

What's New?

• A brand-new front door — hero, portfolio, and team, all shipping today as v1.0.0.9.

• Three-mode theming (light / dark / vibrant) with the two-tier token system — every surface, and every logo, is mode-aware.

• A living hero — "From mindset to grindset, we back companies that apply learning to transform ___", where the last word cycles every 0.9s through the categories we invest in, wrapped in the brand gradient, with an SVG brush-stroke marker under the key words.

• A mode-aware logo marquee — real company trademarks that swap between a dark-ink version (light mode) and a light version (dark/vibrant), normalized into a uniform envelope so a wide wordmark and a square mark read about the same size.

• The full portfolio — 60+ companies rendered from the source data with a Table ⇄ Wall toggle, filters, and vertical tags (HealthTech, LegalTech, EarlyChildhood). Proprietary fields are stripped at build time and never ship.

• A team page with three bio depths (Short / Expanded / CV) you can toggle.

Built by mastering agentic engineering

This rebuild leaned hard into agent orchestration — not as a novelty, but as the fastest correct path:

• ~70 parallel agents each hunted one portfolio company's logo, pulling the most recent full trademark straight off the company's site. When a domain was dead, parked, or rebranded, they recovered the asset through the Wayback Machine and still-live CDNs — so we have real logos even for companies whose sites no longer exist.

• Live rebrands were reconciled along the way (GalileoXP → Kubrio, College Consortium → Acadeum, and more), with a durable rename map so the data regenerates correctly from source.

• The design system, portfolio pipeline, and logo wiring were all built and verified in the loop, following the Astro Knots conventions end to end.

Under the hood

• Palette from the source: LearnStart's light/dark identity was recovered by reading Learn Capital's own compiled CSS — cream, navy, forest, lime — then paired per mode so the gradient stays legible in both.

• Two-tier tokens: named tokens (--color__lime, --color__navy) feed semantic tokens (--color-primary, --grad-*) that re-point per mode. Components only ever read the semantic tier.

• Logo staging: every scraped asset lives in a gitignored .logo-scrape/ staging area (append-only — nothing is ever deleted), and approved trademarks are promoted into public/portfolio-logos/ with role-prefixed names (trademark--light__ / trademark--dark__).

What's Next (why it's .9)

There's more tomorrow — that's the whole reason this ships as 1.0.0.9 and not 1.0.1.0. The .9 is the "epoch" digit doing its job: enough is real and live to stand on its own, with the next pass already queued.

• Wire the remaining ready trademarks into the carousel + wall (Tier S/A first).

• Resolve the "figure out" logo set (Avela, CreatorUP, Vinco, Skillprint, Kubrio, Podium Education).

• Continue per-mode logo QA and optical sizing.

Tracking doc: context-v/plans/Wire-Portfolio-Trademarks-Into-Logo-Surfaces.md.