mozart // entry offer · architecture one-pager · draft v0.1
Take Andrej Karpathy's LLM Wiki method (a knowledge base an AI builds and maintains for you, not one you maintain and occasionally ask about) and turn the single-player laptop version into a hosted, multi-tenant, MCP-accessible company brain on Cloudflare. Mozart installs it, caters the structure to the business, wires it to their tools, and maintains it over time.
Every workflow Mozart builds afterward needs a source of truth to read from and write to. Install the brain first and every later build gets faster and cheaper.
Low-risk, high-trust first project. "Let us build your company brain" is easier to say yes to than a big automation. It opens the door to the workflow revenue behind it.
Once a company's knowledge lives in a brain Mozart installed and maintains, Mozart is wired into the operating core. That is a moat, not a one-off deliverable.
Everything the company ever feeds in, untouched: call transcripts, docs, PDFs, Slack threads, emails, web clippings, meeting notes, screenshots, raw thoughts. The single source of truth of what has been seen. Stored in R2, catalogued in D1.
Encyclopedia-style markdown the AI writes and keeps consistent: entity pages (people, accounts, products), concept pages, summaries, comparisons, an INDEX.md, and backlinks between everything. The AI creates pages, updates them when new raw arrives, fixes cross-references, and resolves contradictions. Rendered from R2/D1, embeddings in Vectorize.
Synthesized responses to queries and scheduled digests, saved as durable records (YYYY-MM-DD-query-slug.md). This is where "what did the brain tell us and when" lives, so answers are traceable, not ephemeral.
Governed by a per-company schema file (Karpathy's CLAUDE.md idea): ingestion rules per source type, index format, backlink + naming conventions, lint rules, and how to resolve conflicting sources. This is the file we cater to each business and the core of the "install."
A source hits inbox/: via MCP, an email address, a webhook, or a tool sync. Nothing is organized yet, it is just safely captured.
The AI reads INDEX.md, extracts entities/concepts, creates or updates wiki pages, writes backlinks, updates the index. Runs on a Cloudflare cron + on-demand.
Health pass: find contradictions, stale pages, duplicates, orphaned links, and concepts referenced but missing. Flag or auto-fix. This is what keeps trust high.
Scheduled rollup of what changed and what is worth a human's eyes. Pushed to Slack/email. Keeps humans in the loop without them doing the filing.
Karpathy's line, and the whole reason this works: humans abandon wikis because upkeep grows faster than value. LLMs do not get bored, do not forget a cross-reference, and touch 15 files in one pass.
| R2 | raw source files + rendered wiki markdown. Cheap, private, no egress fees. |
| D1 | the spine: source catalog, the page/link graph, the intake queue, per-tenant metadata, audit log. (Structured retrieval only, not semantic.) |
| Vectorize | embeddings for semantic search. This is the piece a wiki-on-D1-alone is missing, and it is what makes "ask the brain anything" actually work. |
| Workers + Workflows | the sync / lint / digest jobs, ingestion connectors, and every automation Mozart builds afterward. Durable, retryable, scheduled by cron. |
| Workers AI / model | embeddings + the read-write reasoning that maintains the wiki. Model-agnostic. |
| MCP server (Worker) | the one front door. Humans and agents (ChatGPT, Claude, a Hundo-style agent) query and write the brain through it. Per-tenant auth + scoping. |
| Access / Turnstile | secure the tenant boundary. Each company's brain is isolated and private, which is the whole reason they will not use Notion for this. |
Dedicated inbox email, webhook URL, and an MCP capture tool so a human or agent can drop anything in from anywhere.
Connectors that sweep the company's tools on a schedule: Gmail, Drive, Slack, calendar, CRM, call recorders. Each becomes raw sources.
A small, safe tool set exposed to the company's team and their agents:
search(query) ask(question) → cited answer capture(source) get_page(id) list_recent(entity) digest(range)Because it is MCP, the same brain answers a person in Claude, a teammate in ChatGPT, and an autonomous agent mid-workflow. That universality is the "why now."
"Install your company second brain." Fixed-scope, fast, catered schema + tool connections.
Maintain, monitor, tune the loop, add sources over time. Recurring revenue, ongoing trust.
Every workflow after this reads/writes the brain. Faster to build, easier to sell, higher ceiling.