Rebuilding a 5-Person Engineering Firm's Memory in Notion

How a five-person engineering firm moved its whole memory out of one person's head and into one searchable system — with Notion, Zapier, and a Microsoft API built from scratch.

The challenge

A five-person mechanical engineering firm. Projects that run one to three years. A team spread across four time zones — one full-time engineer overseas, two part-time engineers in different states, a drafter, and senior engineering capacity. They are almost never in the same room, often not even the same time zone.

That structure created one specific problem: the knowledge of record lived in one person's head and across five disconnected tools. Until recently, the only source of truth was "ask Tyler."

• Meeting recordings sat in SharePoint, OneDrive and raw meeting links — permission gaps meant team members sometimes couldn't open them.
• Plaud AI produced transcripts that landed in email, disconnected from the job.
• Client decisions were buried in ten-plus-message Outlook threads that quietly changed requirements halfway through.
• Action items showed up in three places at once — the meeting, the AI summary and a client email — and none of them agreed.
• A failed DIY fix: six hours building a Zapier automation that half-worked. Data came in, but titles and attendees were missing, there was no task tracking, and each Outlook test took over an hour to run.

"I just need another brain. I need a brain that doesn't take days off, that doesn't forget where things are, when a meeting happened, or anything like that." — Tyler Wilson

Solution: one hub, four pipelines

The build has two halves — a Notion workspace the team operates in, and an automation layer that fills it without anyone copy-pasting. The owner set three requirements: make Notion the central hub without forcing the team off SharePoint or Outlook; pull transcripts, summaries and action items into one structured place; and keep large recordings where they live while staying traceable back to source. All three shipped.

1. The Notion workspace.

   Built around how a small async team actually works, not around departments. The core is a set of linked databases — a CRM (companies, contacts, deals), a Meetings database, a Tasks database and a documentation hub for SOPs. Related, not siloed: a company links to its contacts, meetings and emails; a deal links to its proposals, tasks and the meetings where it was discussed. One Meetings table with a type classifier — instead of a separate table per meeting kind — means a new source can be added later without restructuring anything. Tasks are grouped by assignee, then status, so each person opens one view and sees only what they own.

2. Meeting pipeline (SharePoint → Notion).

   When a recording lands in the SharePoint meetings folder, the workflow pulls the transcript out of the raw caption file, generates a structured summary, extracts action items, and writes the whole thing into the Meetings database — title, date, attendees, summary — with each task created and linked back. A filter checks both folder path and file type, so only transcript files enter and subscription noise never does. A code step extracts the readable transcript before any AI touches it, and the meeting title is built from who actually spoke, not the filename. Raw recordings stay in SharePoint; Notion holds the summary, the metadata and a traceable link back.

3. Screen-recording pipeline (OneDrive → AssemblyAI → Notion).

   The most architecturally involved workflow. A webhook catches the recording; a Zapier Tables lookup checks whether it has been seen before. The audio is sent to AssemblyAI, and because transcription is asynchronous, a deliberate delay waits for the job before a second call retrieves the finished transcript. From there it follows the same pattern — AI summary, write to the Meetings database, loop the tasks into Notion. Using a dedicated transcription service rather than the recording platform's own is what makes long screen recordings possible at all.

4. Email pipeline (Outlook → Microsoft Graph → Notion).

   Long client threads were where decisions went to die. Filing a client email into its job folder is the trigger — no approval step, no pasting. The system reads the thread, formats and cleans it, generates a summary, creates an Email record tagged to the job, and produces the action items as linked tasks. Because Outlook offers no clean "new email in a specific job" trigger, the workflow makes a live Microsoft Graph API call to list every subfolder under Jobs, then filters by parent folder ID to resolve the right one. A formatter converts the email to clean Markdown, and a small code step strips characters that would otherwise make Notion's write silently fail. It was also rebuilt for cost — consolidated AI steps and a cheaper model took it from eleven steps down to eight.

The impact

The team now has one place to look. Every meeting, every flagged client email and every task generated from them flows into a single Notion workspace automatically. Any team member can open a job and see what was decided, when, by whom, and what's still open — without messaging the owner. One real job ran fully through the new system and became the firm's best-documented project to date, including a recorded wrap-up that is now a reusable knowledge base for future jobs.

Why this approach works

• Meet the team where they already are. Nobody had to leave SharePoint, Outlook or Plaud. The system wrapped around the existing tools instead of replacing them — filing an email the way you already file emails is the trigger. Adoption cost stayed low because the only new thing to learn was Notion.
• Build from composable parts. Four separate workflows, a dedupe gate and the Notion workspace are distinct pieces with clean joins. When the file-size ceiling appeared, the transcription engine was swapped and the Azure bridge added without touching the other three pipelines. A monolithic build would have meant a rebuild.
• Engineer for the failure modes, not the demo. Most of the real work was in the unglamorous parts — deduplicating duplicate webhooks, stripping characters that break Notion's parser, resolving the right job folder through a raw Graph API call, routing around a vendor's hard size limit. The system works on bad days, not just good ones — the only kind of reliability a five-person team can depend on.

Conclusion

The firm's combined memory moved out of one person's head and into a system the whole team can query — across time zones, without changing the tools they already use.