📖 StoryLine — the storybook studio that lives in your Messages

Text +1 (424) 394-5422 and your kid becomes the hero of a hand-painted storybook — written, illustrated, sold, and delivered by AI agents, entirely inside iMessage. 💙

No app. No signup. No human employees. A parent rambles about their kid; minutes later, illustrated pages are streaming into the thread.

Proven on strangers: during the hackathon, two organic customers texted the number and were served end-to-end — brief → book → in-thread checkout → fulfillment — with zero human involvement. One 👎'd a page, got it repainted in under a minute, and ❤️'d the result.

---

✨ Watch it happen

1. 💬 A parent texts the studio — "my daughter Maya is 4, obsessed with excavators, does everything with our cat Biscuit…"
2. 🧠 Agents extract the brief — hero, age, sidekicks, occasion — via open-weight GLiNER2 on Pioneer
3. ✍️ The story is written, 🎨 the hero is painted — a character sheet locks the kid's look, and every page is generated conditioned on that sheet: same curls, same boots, every single page
4. 📲 Pages stream into the thread as they dry — real iMessage media, arriving one by one like the studio is painting live (because it is)
5. 👎 Tapbacks are the UI — thumbs-down any page and the studio repaints it; ❤️ approves. Reactions are the product interface
6. 💳 Checkout in the thread — $5 keepsake edition, confirmed with a confetti screen effect 🎉
7. 🖨️ The keepsake ships instantly — a print-ready edition, with every order's masterfile preserved in its own paused Superserve VM

🤖 A company, not a demo

• The agents run everything: intake, writing, illustration, art direction (via customer tapbacks), fulfillment, payments, and customer delight — watch it live on the ops dashboard (job rail, press log, revenue meter)
• It survives its vendors: every external dependency degrades gracefully, and the dashboard logs which path served each request — the studio never stalls mid-order
• Unit economics that actually work: ~$0.25 of inference per book against a $5 price → ~95% gross margin 📈, zero marginal labor, in a channel parents already live in. Personalized kids' books are a proven eight-figure market (Wonderbly)

---

🏆 Sponsor tracks — how we used each one

Sponsor	At a glance
💬 Linq	The storefront itself: real iMessage number, pages streamed as media, tapbacks as the product UI (👎 = repaint), screen effects, HMAC webhooks, Agent Pay wired — two strangers became customers by texting it
📦 Superserve	One VM per order as fulfillment architecture: masterfile built and integrity-checked inside, then paused with full state for future reprints — ran on every real order
🧪 Terac	Two verified gen-pop waves that changed the product twice (86% → vector default, 71% → plainer openings), with bot votes quarantined by submission tracking
🐛 Replay	Three explore→fix→verify rounds: 13 real bugs fixed (incl. data corruption), final pass clean, per-finding correctness audit in REPLAY.md, 3 false positives reported
🧠 Pioneer	Open-weight GLiNER2 extracts every customer brief in production — schema-driven NER on rambling parent texts, sub-second, verified on real customers

💬 Linq — the product is an iMessage business

Linq isn't a notification channel here; it's the storefront, the factory window, and the checkout counter.

• A real iMessage number is the entire business: +1 (424) 394-5422. Two strangers became customers during the hackathon by doing nothing but texting it
• Pages ship as media parts, streamed as they're painted — the customer watches their book being made, message by message
• Tapbacks are the product's UI: a 👎 on any page fires a reaction.added webhook and the studio repaints that page; ❤️ triggers a thank-you. Art direction with zero app surface — verified live by a real customer
• Screen effects as brand moments: happy_birthday/sparkles on delivery, confetti on payment confirmation
• Deep webhook integration: Standard Webhooks HMAC verification, pinned payload version (2026-02-03), instant-ack with webhook-id deduplication that correctly honors Linq's retry semantics
• Agent Pay ready: payment_requests integration is wired for in-thread Apple Pay the moment the Stripe connect completes
• Works for email-handle iMessage customers and phone numbers alike — sender identity treated as opaque handles throughout

📦 Superserve — one paused VM per order is the fulfillment architecture

• Every order's masterfile (keepsake HTML + assets + metadata) is written into its own Superserve sandbox; integrity checks and archival run inside the VM
• The sandbox is then paused with full state preserved — Superserve's signature capability used as product architecture: reprints and repaints resume the exact same machine, forever
• This ran for every real order, including both organic customers — fulfillment.json in each order dir records the sandbox ID and integrity report
• Bonus: the whole company runs from an old MacBook Air — the heavy lifting lives in Superserve's cloud, not our hardware

🧪 Terac — human data that changed the product twice (and survived bot contamination)

• Two recruited gen-pop waves through the full opportunities API lifecycle: project → screening question → activity task pointing at our /study A/B page → launch (plus a stop-and-relaunch when our URL rotated — the whole lifecycle, programmatic)
• Participants compared a matched illustration pair (same scene, watercolor vs. vector) and two story openings, tracked per-participant via Terac's submissionId URL params with completion callbacks
• The data shipped two product changes: 86% of verified participants chose the vector style → it became the studio's default; 71% preferred plain openings → the story prompt was updated. Both changes are cited in the code
• The attribution save: anonymous traffic (including an autonomous QA crawler) voted 7–0 the opposite direction of verified humans — without Terac submission tracking, our study's conclusion would have silently flipped. Verified and anonymous tallies are reported separately at /api/study/vote

🐛 Replay — three explore→fix→verify rounds to a clean report

• 21 filings → 13 real bugs found and fixed, spanning build config (dev React served to users), performance (polling cadence, image sizing, a request waterfall), real data corruption (an API field returning [object Object]), and UX robustness (votes silently lost on failed saves)
• Final verification pass: clean — zero findings on the production build
• REPLAY.md: a finding-by-finding correctness audit answering "was the QA right?" — 13 of 14 distinct findings correct (93%), each with verdict, evidence, and fix
• 3 false positives identified and reported (contrast complaints against Cloudflare's tunnel interstitial pages, not our app)
• Documented a novel hazard: the QA crawler voted in our live Terac study — caught and quarantined by attribution

🧠 Pioneer — an open-weight SLM does the core NLU in production

• Every customer brief is extracted by GLiNER2 (fastino/gliner2-large-v1) on Pioneer's native /inference endpoint — schema-driven entity extraction (child name, age, interests, companions, occasion, setting) from rambling parent texts
• Not a demo path: both organic customers' briefs were parsed by GLiNER2 live, logged as served by Pioneer GLiNER2 on the ops dashboard
• Sub-second, purpose-built extraction where a generalist LLM would be overkill — exactly what a small open-weight NER model is for

---

🚀 Run it

pnpm install
cp .env.example .env        # fill in provider keys
pnpm build && pnpm start    # production server on :3000
cloudflared tunnel --url http://localhost:3000
pnpm exec tsx --env-file=.env scripts/linq-subscribe.ts   # point Linq at your tunnel

Then text the number. 📱 Pitch deck: open slides.html. Judge-friendly stops: / (ops dashboard) · /book/<id> (a finished book) · /study (the Terac A/B page).

Under the hood: FLUX.2-pro paints with reference-image character conditioning; DeepSeek-V4-Flash writes the stories.