Vela — autonomous conversion-rate optimization, driven by AI agents

Built for the Unaite × Y Combinator hackathon.

🔗 Live app: lead.159.69.41.115.sslip.io (https://cal.com/youssefb/test-easycro)**

Vela helps companies that own SaaS products ship high-converting landing pages, onboarding flows, and paywalls — automatically. You connect your GitHub repo, an agent orchestrator reads your actual codebase, drafts real code changes, ships them as A/B test variants behind a feature flag, measures conversion, and iterates.

It's a continuous conversion-rate-optimization (CRO) loop, run by AI agents on the customer's own code — every change is a reviewable pull request, never a black box.

Working name in the repo is lead; the product is Vela.

---

The core loop

        ┌──────────────────────────────────────────────────────────┐
        ▼                                                            │
  ① Analyze repo → ② Find opportunity → ③ Agent drafts variant(s)   │
        (audit)        (per surface)        (real code diff)         │
                                                  │                  │
                                                  ▼                  │
  ⑥ Decide  ◀── ⑤ Collect data ◀── ④ Deploy as A/B test            │
   (ship / iterate / drop)   (significance)   (PR + flag)            │
        │                                                            │
        └──────────────► learnings feed the next opportunity ────────┘

1. Connect — log in with GitHub OAuth; Vela reads your repo over the GitHub API (no clone).
2. Audit — a discovery agent (Claude Opus 4.8, agentic file-read loop) locates your landing / onboarding / paywall surfaces; per-surface analyzer agents score them and propose changes.
3. Launch — the launch agent picks one high-leverage copy test and opens a reviewable PR that tags the element with our SDK data-attributes (control copy stays the visible default). The PR body embeds a rendered Before/After preview image.
4. Track — our own lightweight SDK (public/sdk.js) captures pageviews, clicks, and conversions, does deterministic variant assignment, and beacons events to our ingest endpoint.
5. Measure & decide — the dashboard computes uplift + confidence via a two-proportion z-test; gated Activate / Ship winner / Abandon controls flip the flag live with no redeploy.

---

Why it's different

• Real code, not a visual overlay. Variants ship as pull requests against the customer's repo. Reviewable diffs, their CI, their deploy.
• One integration. A single SDK snippet owns both variant delivery and conversion tracking — we own assignment and the statistics.
• Review-gated autonomy. The agent finds opportunities and drafts variants automatically, but a human approves before anything goes live. An autonomy dial loosens this over time.
• Live flag flipping. GET /api/experiments/active serves only RUNNING experiments, so a test goes live (and stops) the instant you flip it — no redeploy.

---

Architecture

Layer	Choice
Framework	Next.js 16 (App Router), TypeScript strict
UI	shadcn/ui + Tailwind v4 — "Warm Precision" design system (oklch tokens, light + dark)
Database	PostgreSQL via Prisma 7 (pg driver adapter)
Auth	Better Auth + GitHub OAuth
Agents	Claude Agent SDK / Anthropic SDK — orchestrator on claude-opus-4-8
Repo writes	GitHub App installation token (RS256 JWT) for PRs; OAuth token for reads
Tracking	First-party SDK (public/sdk.js) → POST /api/ingest
Deployment	Self-hosted via Coolify (live here)

What's actually built (not mocked)

• ✅ GitHub OAuth + repo connect → persisted Project
• ✅ Real codebase audit — discovery + per-surface analyzer agents, streamed as NDJSON, persisted as a ProjectAudit
• ✅ Real A/B launch — agent opens a PR with embedded Before/After preview; experiment + variants minted first to embed their ids
• ✅ Real visitor tracking — first-party SDK, CORS ingest, impression/conversion rollups, live dashboard band
• ✅ Real stats + decision controls — two-proportion z-test uplift/confidence; gated Activate / Ship winner / Abandon
• ✅ Setup agent — injects the tracking snippet + conversion markers via a reviewable PR

---

Try it live — no setup, no clone

Vela is already deployed and running. There's nothing to install — just open it in your browser and test the real product:

👉 lead.159.69.41.115.sslip.io

Then:

1. Log in with GitHub and connect a repo.
2. Watch the audit agent read the codebase and surface real opportunities.
3. Launch an A/B test — a reviewable pull request is opened on the repo, with an embedded Before/After preview.
4. See conversions and uplift / confidence flow into the dashboard.

Prefer a guided walkthrough? Book a 1:1 demo →

---

Repo map

app/
  api/            # audit, experiments, github, ingest, projects, setup-sdk, auth
  dashboard/      # surfaces, experiments, insights, agent, settings, overview
  onboarding/     # connect → repo → audit → report flow
lib/
  agents/         # audit.ts, launch-experiment.ts, setup-sdk.ts (Claude Agent SDK)
  github.ts       # repo reads (REST) + multi-file commit / PR helpers
  github-app.ts   # GitHub App JWT → installation token (PR writes)
  metrics.ts      # event aggregation
  stats.ts        # two-proportion z-test
  preview-svg.ts  # Before/After preview rendering
public/sdk.js     # first-party tracking + variant-assignment SDK
docs/product.md   # full product & dashboard spec

---

Built with Claude Code for the Unaite × Y Combinator hackathon.