An integrated AI ops stack for a VC firm.
BHI had every tool a venture firm needs — Attio for CRM, Notion for diligence, Granola for meeting notes, Box for files, Specter for company data, Gmail and Slack for everything else. None of them talked to each other. In roughly one month, solo, I built the connective tissue: an AI-native operating layer that turned six disconnected systems into one queryable workspace, with eight purpose-built tools the deal team uses today.
- Role
- Sole engineer · stakeholder discovery → architecture → build → roll-out · shipped end-to-end in ~1 month
- Year
- 2025
- Status
- In production
Problem
The data wasn't missing. The SOPs weren't missing. What was missing was a single place a partner could open and see everything we knew about a given company. Six tabs to answer one question, and half the time the version in Attio was three weeks behind the version in someone's Granola transcript.
The brief I gave myself was deliberately small: don't replace the tools, connect them. Pick a hub (Attio), pick a file home (Box, after a small comparison against OneDrive and Google Drive on MCP integration completeness), and make Claude the interface that pulls across all of them in one conversation.
Layer 1 — MCP across the stack
Connected Claude to seven of BHI's existing tools via the Model Context Protocol: Attio, Notion, Granola, Gmail, Box, Google Calendar, Slack, and Specter. The result is a single conversational surface where an partner can say "pull everything we know about Conflixis" and Claude searches Attio, Notion, Granola, and Gmail simultaneously, then posts a Slack summary or writes a note back to Attio in the same turn.
The architecture decision that made the rest work: Attio is the hub. Every other system either feeds into it (Granola notes, Box folder URLs, Notion DD pages) or reads from it (Slack briefs, Pre-Call cheat sheets). Two clicks from any company record to anything written about it.
Layer 2 — Eight custom Claude Projects
MCP gives Claude the keys; the team needed purpose-built tools on top of it. I shipped eight Projects (custom instructions + MCP connections, shared with the whole firm — no setup required from the partner):
- Pre-Call Brief — 4–5 sentence cheat sheet before every meeting: last contact, key questions, funding status.
- Company Research — structured VC snapshot from internal sources + web.
- Everything About a Company — full data dump: basics, team, financials, traction, interactions.
- Internal Scorecard— scores companies on BHI's official scorecard using internal data, with citations.
- External Scorecard— same scorecard from public research only ("should we take the meeting?").
- Sentiment Analyzer — analyses Granola call transcripts for engagement, transparency, urgency.
- Communication Coach — analyses your side of a call and coaches on questions, listening, rapport.
- Dealflow Router — takes mixed meeting notes and splits them into the correct Attio company timelines.
Layer 3 — Cowork (scheduled automations)
The Projects are pull-based. Cowork is push-based — Claude driving a computer on a schedule, doing the small filing tasks nobody actually wants to do.
- Scan Gmail for unfiled deal attachments → Slack reminder. Weekly.
- Check Attio for missing Box / Notion links → Slack hygiene checklist. Friday.
- Flag stale pipeline companies. Configurable cadence.
- Create a Notion DD page from template → push the URL to Attio. On demand.
- Create a Box folder for a new company → link it to the Attio record. On demand.
- Push Granola notes into Attio in the right structure. On demand.
The point of the layer wasn't to replace SOPs — it was to remove the friction that makes people skip them.
Layer 4 — Custom code (server-side)
Two pieces needed real code rather than no-code:
- Portfolio news monitor— a Python script that checks news for portfolio companies on a schedule, posts weekly digests to Slack, and is fully configurable through a Notion table (add / remove companies, change frequency, switch the Gemini model). Runs on GitHub Actions' free tier — $0 in infrastructure.
- Portfolio dashboard — Cowork-driven exploration that reads PDFs and Excel models across the portfolio, extracts revenue / growth / margins / team size, and generates comparison visualisations. Started as a one-off prompt and became a recurring tool.
The pattern
Four layers, one idea: connective tissue. The tools existed. The data existed. The SOPs existed. Each layer above turns a manual step into something that happens automatically — without changing how anyone on the team actually works.
- End-to-end build time, solo
- ~1 month
- Custom Claude Projects in production
- 8
- Scheduled Cowork automations
- 6+
Four layers, eight Projects, six automations, one custom Python service running for $0 on GitHub Actions — built solo in roughly a month, including the stakeholder discovery that decided what to build in the first place. The investment memo builder lives within this same stack and gets its own deep-dive.