Three.js · WebGL 2
Custom GLSL shaders, cursor-aware geometry, scroll-linked motion. 60 fps desktop, 30 fps mobile fallback, JSON-driven scene config.
01 · The problem
Pretty brand sites don't sell. Sales engines aren't pretty.
Most premium brand sites are beautiful and inert — they win the click and then do nothing. Meanwhile, the sales stack that would convert that click (content marketing, cold outreach, CRM, A/B learning) typically requires a marketer, a copywriter, a sales-ops engineer, and a growth PM — four hires before the first email goes out.
BrandiVibe proves a solo operator can ship both layers — a 60-fps WebGL front-end, and the autonomous sales engine that feeds it qualified pipeline — as one system.
02 · The approach
A front-end that earns the trust, and an AI Sales Brain that converts it.
Layer 1 — Experience. Three.js with custom shaders, cursor-aware geometry, scroll-linked motion. JSON-driven scene config so the brand edits copy and colour without touching render code. 60 fps desktop target; 30 fps graceful fallback on mid-range mobile.
Layer 2 — The AI Sales Brain. A GPT-4o-driven autonomous system running on Koyeb with state persisted to GitHub. It writes one long-form SEO article every day, scans four startup feeds hourly for ICP-fit prospects, researches each one (scrapes homepage/about/pricing/contact, extracts personalization hooks), and runs a warmup-curve cold-email sequence via Resend. A weekly digest scores every email by angle, industry, and subject style — then proposes and auto-applies A/B winners. Zero daily human input.
03 · Architecture
How it fits together.
GPT-4o · daily
1,800–2,500-word SEO article per day. 10 content angles rotating (ROI, conversion psychology, case studies…). Pexels hero images, 40+ banned AI-phrase filter for E-E-A-T, auto-publishes to /journal.
4 feeds · hourly
Scans TechCrunch, Product Hunt, Hacker News Show HN, and BetaList every hour. GPT-4o extracts company, domain, funding, industry. Filters out mega-brands and sites already using premium design (Three.js, Spline, Framer, GSAP) to keep only ICP fit.
Scrape + GPT-4o hooks
Scrapes each prospect's homepage, about, team, pricing, contact. Tech-stack + design-quality score. GPT-4o extracts personalization hooks — specific weaknesses, decision-maker name, budget estimate. Email found via scrape then pattern fallback with MX verification.
Resend · warmup · timing
Personalized 4-touch sequence per prospect. Warmup ramps 5 → 50 emails/day over 30 days. Per-domain throttle (1 email / 3 days), send window Tue/Wed/Thu 13:00 UTC. Smart timing: clicked = 2-day follow-up, opened = 3 days, no-open = 5 days with a new subject.
Webhooks · CRM
Resend webhook records opens, clicks, bounces, replies. Tracked /api/brain/book link redirects to Cal.com and marks booking intent. Inbound replies auto-pause sequences. Everything lives in brain.json, synced to GitHub.
A/B · auto-apply
Monday 14:00 UTC digest scores every email by angle, industry, and subject style. Identifies winners and losers. Proposes A/B tests between top-performing patterns. Auto-applies winners to the planner's prompt when BRAIN_AUTONOMY=true. 60-day override expiry so it re-tests and doesn't freeze on stale truth.
Self-healing · resumable
Six-hour health check emails mjrifat54@gmail.com on failure. DailyRun ledger lets the pipeline resume from its last checkpoint if Koyeb kills mid-tick. State synced to GitHub survives every redeploy.
04 · The work beneath the surface
Non-obvious decisions.
- Frame-budget discipline — every new scene element has to fit within a fixed frame budget. Nothing ships that drops the experience under 45 fps on a 2019-era laptop.
- Deliverability engineering — the warmup curve, per-domain throttle, and narrow send window aren't optional — they're what keeps the from-domain off spam filters and makes the engine sustainable past month one.
- ICP filtering that refuses customers — Lead Radar actively rejects prospects that already ship premium design (Three.js / Spline / Framer / GSAP detected on their site). They're not the ICP, so they're noise — even if the brain could get a reply.
- Self-experimentation with guardrails — the brain proposes its own A/B tests and applies winners, but overrides expire after 60 days. That stops it from freezing on a pattern that stopped working six months ago.
- State as version-controlled artifact — brain.json sits in a GitHub repo. Every run commits a diff. The entire history of every prospect, email, and learning decision is auditable — and Koyeb redeploys can't wipe it.
05 · Stack
What it's built with.
Front-end
Three.js · GLSL shaders · WebGL 2 · JSON-driven scene config
Intelligence
GPT-4o (planner, article writer, lead extractor, personalizer)
Sourcing
TechCrunch RSS · Product Hunt · Hacker News Show HN · BetaList
Delivery
Resend (warmup curve, webhooks for open/click/bounce/reply)
Infrastructure
Koyeb cron workers · brain.json on GitHub · Pexels media · Cal.com booking redirect
Schedule
Daily 13:00 UTC · hourly sourcing · 6-hour health check · Mon 14:00 UTC learning digest
06 · Takeaway
Why this one matters.
BrandiVibe is both the storefront and the sales engine. The WebGL front-end earns the trust; the AI Sales Brain behind it converts attention into pipeline — writing daily SEO articles, sourcing ICP-fit prospects from four startup feeds, researching them with GPT-4o, running a four-touch warmup-curve cold-email sequence, and auto-applying its own A/B winners every week. One operator, two layers, zero daily human input. Most companies form a team to build half of this.
Want something like this built for your company?
I'm currently available for select engagements.