Build with Gemini XPRIZE — Devpost Submission Draft (LumiBase)

Nháp soạn ngày 2026-06-13. Hạng mục: Small Business Services. ⚠️ Các mục đánh dấu [TODO] cần bằng chứng thật từ founder trước khi submit lần cuối (ban giám khảo có verify repo, log, doanh thu).

Lưu ý về ngôn ngữ: phần narrative và các trường "About the project" bên dưới là văn bản sẽ nộp nguyên văn cho Devpost, nên được giữ nguyên tiếng Anh. Chỉ phần khung, hướng dẫn và checklist được dịch — đó là phần bạn cần đọc để biết mình phải chuẩn bị gì.

Tên project

LumiBase — The Content Operating System

Elevator pitch (≤200 ký tự)

AI agents run your content operations — writing, translating, SEO, cleanup — against declarative SLOs, with earned autonomy and full provenance. Humans set intent; agents do the work.

Hạng mục

Small Business Services

Built with (tags)

gemini, google-cloud, typescript, hono, cloudflare-workers, postgresql, drizzle-orm, turborepo, react, docker, zod

Narrative (500–1000 từ) — nộp nguyên văn tiếng Anh

How AI runs this business, day to day

LumiBase is a Content Operating System: a runtime where AI agents operate content for small businesses while humans set intent, taste, and accountability. A traditional CMS is a tool a human clicks through. LumiBase inverts that. A customer declares the desired state of their content as an SLO — "every published product has at least one image, a 50–200 word description, and Vietnamese + English translations" — and a control loop of governed agents converges the content toward it, continuously.

Day to day, AI does the operational labor of a content team. A reconciler scans every site against its SLOs and detects drift: products missing translations, stale blog posts, broken links, items without images. Each drift becomes a goal. A planner agent decomposes goals into sub-goals and delegates to a newsroom of narrow-scoped role agents — Writer, Translator, Taxonomist, SEO agent — each holding only the capabilities its role needs. A Reviewer agent fact-checks low-risk output; high-risk output escalates to a human. Gemini powers the reasoning core of this harness: goal decomposition, content generation, and the LLM-judge evaluators that enforce each tenant's editorial constitution.

AI is not a feature bolted onto the product; it is also how we build and run the company. The codebase ships agent-setup contracts (CLAUDE.md, AGENTS.md, .cursorrules, a machine-readable prompt.md) so coding agents implement features under the same non-negotiable rules human contributors follow. Support, onboarding flows, changelog drafting, and documentation are produced agent-first and reviewed by a human.

What humans do versus what AI does

Humans hold three jobs the system is explicitly designed around. They set intent: SLOs, goals, budgets. They encode taste: each tenant writes a versioned, hashed "constitution" — brand voice, legal boundaries, quality bars — as rule-DSL and LLM-judge evaluators; an artifact that fails the constitution never publishes, at any autonomy level. And they keep accountability: every agent action is revertable, every revision records which agent, run, and model produced it, what sources it used, which constitution hash it was judged against, and who approved it.

Everything else — data entry, translation, tagging, SEO fixes, description writing, content cleanup — is agent labor. Crucially, autonomy is earned, not granted. Each (site, agent, capability) tuple sits on a five-level trust ladder: Shadow → Propose → Co-sign → Veto-window → Autopilot. Promotion is data-driven (N consecutive passing runs, approval rate above threshold, zero incidents); demotion is automatic on any incident. At the veto-window level, agents execute into staging and commit after a timeout unless a human vetoes — humans move from pre-approval to exception handling, which is what makes one person able to govern thousands of agent actions a day. A four-scope kill switch is always available.

Jobs and economic opportunity beyond the founding team

LumiBase targets the small businesses that can least afford content operations: the local retailer who needs a bilingual product catalog, the agency serving ten such retailers, the solo founder who cannot hire an editor, a translator, and an SEO specialist. By replacing operational content labor, it lets these businesses compete with the content quality of much larger players.

It also creates new, higher-value work. The "editor" role shifts to constitution author — a person who encodes editorial judgment once and leverages it across every agent action thereafter, a 1→N multiplier instead of 1→1 editing. Agencies on LumiBase can serve more clients per head and sell governance (constitutions, SLO design, exception review) as a recurring service. Because the core is Apache-2.0-licensed and the skill registry is exposed as a standard MCP server, third-party developers can build and sell agents, skills, and extensions on top of the platform — an ecosystem of small AI-native service businesses. [TODO: add actual numbers — pilot customers, agencies onboarded, marketplace contributors]

The story of building this way

LumiBase began as an edge-native headless CMS. Partway through, we confronted an honest question: if agents are about to do most content work, why build another UI for humans to click? We refactored the product around the answer (v0.5.0): the primary API became POST /agent/goals rather than POST /items; the UI became Mission Control — exception inbox, trust ledger, kill switch — rather than an editing surface. Every new feature must answer "how does an agent call it?" before "how does the UI show it?"

Building a business this way means the founding team stays tiny while the operating capacity scales with compute. The hard problems turned out not to be generation quality but governance: how to let agents act without letting them act wrongly, and how to make silence-means-consent safe. The trust ledger, the constitution gate, and provenance-first revisions are our answers — and they are also why a small business can trust an AI-operated system with its public voice.

[Mục tiêu số từ: 500–1000. Hiện tại ≈ 760]

Bằng chứng doanh thu

[TODO — phải là số thật:]

Giá: core open-source Apache-2.0 (free) · Hobby $29/tháng managed hosting · Enterprise $99/tháng (SSO, hỗ trợ riêng)
~~⚠️ Kiểm tra license trước khi nộp.~~ Đã giải quyết. License là Apache-2.0, không phải MIT, và đây là việc đã chốt chứ không còn mở: CHANGELOG.md ghi việc đổi license ở mục [0.23.0] (2026-07-14) — "Project license updated to the Apache License, Version 2.0 (from MIT), effective this release", với v0.22.0 là bản MIT cuối cùng. Repo đang ở 0.26.0, LICENSE ở gốc là Apache-2.0, và cả bốn package được publish đều khai "Apache-2.0". Các chỗ nhắc MIT trong bản nháp này đã được sửa.
Export dashboard Stripe / sao kê ngân hàng / P&L đơn giản (template: https://docs.google.com/spreadsheets/d/1pAJrEMo7_QID6V62sA4C8XwGBHkxDTVX3wtYNE2fulI/edit)
Mã số doanh nghiệp nếu có

Chi phí (trong kỳ hackathon)

[TODO — phải khai kể cả khi bằng 0:]

Chi cho marketing & thu hút khách hàng: $___
Hạ tầng (Cloudflare, Google Cloud/Gemini API): $___
Khác: $___

Bằng chứng sản phẩm (AI đang chạy trên production)

[TODO — export từ hệ thống đang chạy:]

Log thực thi của agent (các bảng agent_runs, agent_goals, agent_evaluations)
Bản ghi sử dụng Gemini API (screenshot billing/quota trong Google Cloud console)
Screenshot Studio Mission Control: exception inbox, trust ledger, các lần reconciler chạy
Response của Provenance API (?provenance=true)

Bằng chứng khách hàng

[TODO — chỉ khách hàng thật:]

Tên / email / điện thoại của khách hàng thật
Testimonial hoặc feedback

Các trường "About the project" của Devpost — nộp nguyên văn tiếng Anh

Inspiration: Content teams at small businesses drown in operational labor — translation, tagging, SEO, cleanup — that AI can now do, but no CMS was built for agents to operate safely.

What it does: Declares content state as SLOs; governed AI agents continuously reconcile content toward them, with earned autonomy (L0–L4), a tenant constitution gating every publish, and full provenance on every revision.

How we built it: Turborepo monorepo; Hono on Cloudflare Workers (edge delivery) with a Docker/Node adapter; PostgreSQL + Drizzle; Zod contracts shared across CMS, Studio, and SDKs; Gemini as the agent reasoning + LLM-judge layer; skill registry exposed as an MCP server.

Challenges: Designing governance that scales — binary HITL doesn't; we built the trust ladder and veto-window (silence-means-consent with absolute rollback).

Accomplishments: v0.5.0 Content OS refactor: reconciliation loop, trust ledger, constitution publish-gate, multi-agent newsroom, provenance-first revisions.

What we learned: The hard problem is not generation, it's accountability. Autonomy must be earned per capability, and every byte needs a lineage.

What's next: Agent marketplace on the MCP surface, more Gemini-powered role agents, managed-hosting growth for SMB customers.