Middleman
An AI-native deal workspace for financial professionals.
- Role
- Co-founder and product lead, Arsun Technologies
- Status
- Paused after validation
- Year
- 2025
- 98k+
- 10k+
- 570+
- 33
- 6+
The problem
A banker running a live deal works across email, spreadsheets, a document repository, a CRM, a calendar, a financial model, a task list, and — increasingly — three or four disconnected AI tools. Nothing knows about anything else. The deal exists only in the head of the person holding it together.
The question Middleman started from: what does this workspace look like if you assume AI and agents from the first line of code, rather than bolting a chat panel onto software designed in 2011?
What I built
One workspace covering the deal end to end — pipeline with custom stages, a vault for uploaded and generated documents, tasks, teams, calendar, and a full email client with AI-assisted replies.
On top of that, a generation layer for the documents the work actually produces: CIMs, teasers, investment memos, management presentations, DCF and LBO models — with citations back to source material so an output can be traced rather than trusted.
And an agent layer with graduated autonomy, from observe-only through execution with human approval, so the software could earn trust incrementally instead of demanding it up front.
Selected details
- Workspace
Deal pipeline and CRM with custom stages, document vault, tasks, teams, calendar, and a Gmail-connected inbox with AI-drafted replies. Mentions, tags, and approval workflows ran across all of it.
- Document generation
A studio for CIM, teaser, investment memo, and management presentation generation, plus DCF and LBO modeling. Every generated claim carried a citation back to the source document.
- Agents
EmailAgent and InboxAgent handled correspondence; a playbook engine encoded firm-specific process; a behavioral replication mode watched how a user worked and proposed actions in that style. Autonomy was a dial, not a switch — observe, suggest, execute-with-approval.
- Retrieval
Hierarchical retrieval over deal documents with Voyage embeddings and reranking, Document AI and OCR for the scanned material that shows up in diligence, and streaming responses over SSE.
- Stack
React, TypeScript, Vite, Tailwind and shadcn on the front end. Firebase throughout — Firestore, Storage, Functions, Scheduler, Hosting. Gmail and Google Calendar OAuth, Stripe for billing, Vertex AI, Gemini and Claude for inference.
Screens
Where it landed
The product got real traction, and not enough of it to justify what it cost to build. That's the honest version.
The thesis was that boutique and middle-market firms would move their core workflow onto something better. They were more reluctant to switch than the thesis assumed — the existing stack is bad, but it's load-bearing, and replacing it mid-deal is a risk nobody wants to own.
What I took from it
- Engineering velocity is not a substitute for distribution. I could build faster than I could find out whether anyone wanted it, which is a worse position than it sounds.
- Switching costs invalidate otherwise compelling workflow improvements. "Better" has to clear a much higher bar than better.
- Validation belongs before the large engineering investment, not after it. The expensive version of this lesson is the one I took.
- A technically impressive product is not automatically a business, and a backlog is not a decision.