SuperBrain

Inspiration

The web is shifting from humans to agents — and publishing without a citation rail is writing for free. When ChatGPT summarizes your page to a user, no click lands on you. No attribution. No payment. Meanwhile, existing AI research tools hand you a
chat answer when what other agents actually need is a typed data record. Senso's cited.md manifesto — every fetch by every agent is a transaction — and Karpathy's "compounding knowledge" thesis gave us the thread: build a node in that new agent-native web, not another SaaS dashboard on top of the old one.

What it does

You type a company name. In under a minute, SuperBrain:

Dispatches 4 parallel TinyFish browse agents across the company website, LinkedIn, X search, and Google News
Writes every extracted fact to Redis Cloud as semantic memory and flags
contradictions between sources
Retrieves grounded context from our curated Senso knowledge base
Hands the raw pages + KB context to Gemini (gemini-3.1-flash-lite-preview) to
synthesize a typed CompanyInsight with real named people, real competitors, and
URL-cited contradictions
Publishes a free-tier post to Ghost Pro with an interactive 3D WebGL knowledge map embedded inline
Publishes a full citeable to cited.md via Senso so downstream agents can cite and micro-pay us for the fetch
Exposes the knowledge as a federated GraphQL subgraph through Wundergraph Cosmo
Gates the deep-research section (contradictions, people, full sources) behind a 1 USDC x402 payment settled by a Coinbase Developer Platform facilitator

The agent narrates itself in a live SSE sidebar the whole time. The report is
simultaneously a Ghost article, a cited.md citeable, a GraphQL subgraph, and a paywalled dashboard — one record, four surfaces.

How we built it

Next.js 16 (App Router) + React 19 + TypeScript for the full-stack app
React Three Fiber + drei for the 3D knowledge map, embedded in Ghost posts via a chromeless /embed/map/:id iframe
Adapter pattern for every sponsor (src/lib/adapters/*.ts) — each adapter has a real integration plus a graceful in-process fallback, so the orchestrator doesn't know
whether it's talking to the real cloud or a mock
Server-Sent Events for live agent status streaming into the sidebar
Apollo Federation v2 subgraph served by a direct graphql-js executor at
/api/graphql, then registered with wgc and composed into the Cosmo federated graph
superbrain
Gemini's structured JSON output with a strict schema + "every contradiction must
cite a URL from the inputs" prompt rule, so the LLM can't hallucinate people or
sources
Senso onboarding skill installed via Shipables, then run end-to-end: 7 folders,
brand kit, 4 content types, 12 KB docs, 9 prompts, 3 live citeables, GEO monitoring
on Mon/Wed/Fri
x402-next's withX402 wrapper + @coinbase/x402 facilitator for the real payment rail
Neural Nexus design system — cyan/violet glassmorphism, Inter + Space Grotesk, glow states — applied end-to-end with a brain-shaped landing background
Deployed as a single Next.js app with graceful fallbacks everywhere — if any
sponsor is down mid-demo, the loop continues on local mocks and we never blank-screen on stage

Challenges we ran into

Senso's public docs point at the wrong base URL — sdk.senso.ai returns 404s for
writes. We decompiled the CLI's bundled JS to find the real endpoint (apiv2.senso.ai/api/v1/org/content-engine/publish) and the real auth header.
TinyFish's sync /run endpoint blocks for 30–60 seconds per URL. Four sequential calls would blow the 3-minute demo budget. We parallelized them and tightened the
per-call timeout to 25 seconds, so worst case 1–2 of 4 cleanly fall back to mocks.
JSX resolves to the SVG element, not THREE.Line. The knowledge map's edges literally weren't rendering. Swapped to stretched cylinder meshes — unambiguous and visually thicker, with opacity driven by edge strength.
engine publish returned "content linkage conflict" when a draft already existed for a prompt. Solved with delete-then-publish-fresh in our adapter.
Type mismatch between @coinbase/x402 and x402/types — both ship a FacilitatorConfig that's structurally compatible but not type-assignable. One as unknown as cast at
the boundary.
Nexla's account was empty and flow setup didn't fit the weekend. Rather than ship a broken claim, we silently dropped Nexla from every user-facing surface and added a
replacement in the narrative.
Keeping 7 sponsor credentials coordinated without leaking any of them into git.
.env is gitignored; pre-commit secret scans ran on every staged diff.

Accomplishments that we're proud of

Seven real sponsor integrations in a weekend, not four real and three mocked —
Redis, Ghost, Senso, Wundergraph Cosmo, TinyFish, CDP + x402, and Gemini all talking to real clouds.
Gemini returns real, correct facts. On a "Notion" test run, unprompted, it named Ivan Zhao, Simon Last, and Akshay Kothari as founders and listed Atlassian
(Confluence/Jira), Microsoft Loop, Obsidian, and Coda as competitors — every one accurate, none hallucinated.
Every run ships all three publishing targets simultaneously — a live Ghost post with a spinning 3D WebGL map, a cited.md citeable, and a federated GraphQL record.
One record, four surfaces.
Graceful-fallback adapter mesh — every sponsor can fail and the demo keeps running.
Cosmo federated graph live and composable with our subgraph published and checked.
Neural Nexus design system applied end-to-end — landing, agent status, report,
embed — with glassmorphism, glow states, and a brain-shaped background.
Full loop in ~35 seconds with all real clouds firing.

What we learned

LLM structured-JSON output with schema + "cite-from-inputs" rules produces usable
data — Gemini didn't hallucinate once across our test runs because the system prompt required URLs from the provided pages for every contradiction.
Parallelism plus strict timeouts is a demo-safety pattern, not a hack. 4 × 60s sequential would have been 4 minutes. 4 × 25s parallel is 25 seconds worst case.
Adapter pattern pays for itself at 3+ sponsors. Swapping mock to real, or real to fallback, should be one file and no touch to the orchestrator.
The agentic web's pricing model is genuinely different. Per-seat SaaS assumes users; agents don't sign up. Per-request micropayment matches the unit of work.
"Graceful fallback" is a feature, not a crutch — it's what separates a demo that ships from one that breaks.
Decompile the CLI when the docs lie. Twice this weekend, reading node_modules/.../dist/cli.js unblocked what the public docs left vague.

What's next for SuperBrain

Autonomous decision loop. Today the orchestrator is a fixed 6-step pipeline. Next: Claude or Gemini decides what to browse next based on what's already found ("the website claims X but LinkedIn is thin — check Crunchbase and the founder's X feed").
TinyFish /run-async with streaming. Today we use the sync endpoint and cap at 25s. Async + SSE polling lets us wait longer for richer extraction without a cold demo
budget.
Real on-chain settlement on base mainnet. CDP facilitator is wired; we need a real CDP-controlled wallet and the end-to-end x402 client flow on a live invoice.
MCP server wrapper. Expose SuperBrain as a tool Claude, Cursor, and Codex can invoke natively — "run SuperBrain on Anthropic, ground it, and cite it back."
Cross-subgraph federation. Combine our superbrain subgraph with public subgraphs (people data, funding data, SEC filings) so downstream agents query one unified
GraphQL supergraph.
Scheduled re-runs. Track a company, receive a diff report when something material
changes (new CRO, sentiment flip, contradiction resolves).
GEO feedback loop. Use Senso's scheduled GEO probes across ChatGPT/Claude/Perplexity/Gemini to see where our published citeables are winning and losing, then tune the prompts + KB automatically. The compounding flywheel Karpathy described — but for a company, not a person.