Owner-led SMB
10 to 200 employees, high operating density, many signals outside clean SaaS tables.
Thesis
Company Brain is the category. Raumdeuter is the version for physical, owner-led operations.
Glean starts from enterprise search. Hebbia starts from research. GBrain starts from founder memory. Raumdeuter starts where ordinary SaaS data models break: workshop photo, message commitment, customer-update gap, billing window, owner judgment.
Operational Radar
Radar Cards with signal, evidence, proposal and approval gate. That is the MVP surface, not a multi-feature bundle.
Operational physics
The lab is hard, fragmented and regulated. That difficulty becomes product substance.
Six differentiation anchors
Not a better Glean. A different category.
Owner-led SMB
What breaks today if the owner has to keep it all in his head?
Physical evidence
A photo can be proof, risk, instruction, customer update, billing basis and story material at once.
Proactive radar
Raumdeuter does not wait for search. It surfaces operational knots before they become damage.
Sovereignty
Raw data stays in customer storage. Raumdeuter stores pointers and narrow evidence slices.
Failure-to-rule
Operational misses become rules, eval cases and cards. The harvest is running, external count claims wait for the table.
Operator-founder
Marc is operator and builder. Raumdeuter comes from daily operating friction, not market slides.
Map
How Raumdeuter sits against Company Brain competitors.
| Dimension | Glean / Hebbia / GBrain | Raumdeuter |
|---|---|---|
| Origin | Enterprise Search, Finance Research, Founder Brain | Real owner-operated business |
| User | Knowledge worker, analyst, founder-builder | Overloaded operating owner |
| Data reality | Docs, apps, meetings, Slack, CRM | Messages, photos, videos, invoices, workshop notes, commitments |
| Main mode | Search, analysis, agent memory | Proactive radar |
| Output | Answer, analysis, agent action | Five card types with evidence and approval |
| Moat | Distribution, data integration, scale | Operator-founder, hard market, real failure loops |
Founder leverage
AI-native Workbench is part of the proof, not the MVP hero.
Marc operates a file-based AI work system today: Claude/Cowork as strategy, coaching and approval layer, Codex as worker layer for research, systematization and execution. The customer surface stays narrow: Operational Radar with five card types.
HandoffsYAML headers, status markers, review gates, clear folder discipline.
Worker LayerCodex handles research, ETL, bulk operations, QA and website builds.
ApprovalExternal impact stays behind Marc and review gates. No auto-publish thesis.
Investor Q&A
Short answers to the first hard questions.
What are you making?
Raumdeuter is the AI Operating System for owner-led SMBs where the owner is the bottleneck. It pulls scattered data into an evidence graph, runs versioned skills and delivers radar cards with evidence and approval gates.
How are you different from GBrain, Glean and Hebbia?
We do not host the full data estate. Raumdeuter is designed as an index on the customer's file system, plus operator memory and an eval library from real operating misses.
Why now?
LLMs are reliable enough for fenced skill tasks, owner-led SMBs face labor and regulatory pressure, and AI-native founder building makes pre-seed leverage realistic without a classic CTO setup.
Why use classic-car restoration as the lab?
Because the market is hard: long project cycles, fragmented communication, physical evidence, emotional customers, parts chains, payments and regulation. If Raumdeuter works there, transfer to other service businesses is plausible.
How does concierge become product?
First stabilize three pilot companies with weekly radar, then productize the multi-tenant substrate, skill manifest and eval log. The concierge MVP is the learning engine, not the end state.