Olympus / Independent applied AI project
AI workflows that
verify their actions.
Olympus connects agents, email, meetings, and knowledge management. I designed it around clear requirements, useful integrations, and evidence that an action actually completed.
Simplified architecture. Different workflows use different checks; an uncertain action stops for inspection.
From scattered information
to useful follow-up.
Meetings, notes, and follow-up work lived across separate tools. Turning that information into usable knowledge meant repeating the same manual work—and checking that nothing important was lost between tools.
I designed a personal system across two Macs that connects meeting capture, email handling, and a Notion/Obsidian knowledge base. Each workflow has a defined source, an owner, allowed actions, and a checkable finish.
Connect the knowledge.
Keep the source.
I built Notion teamspaces and automations and directed AI-assisted integration work linking capture, knowledge, and follow-up. Source references remain attached to processed information so I can check where a statement came from.
The design separates information storage from permission to act: a saved note is context, not a command to an agent. Existing records are checked before new ones are created, and destination updates are read back after they save.
I also worked on backup coverage and sample recovery so the system could be maintained beyond a successful demo.
Design the failure path.
Uncertain API results
Record the attempt before dispatch. When an acknowledgement is lost or invalid, preserve the unresolved state and check the destination before trying another change.
Incomplete sources
Keep a failed read distinct from an empty result. Check source identity and revision before processing, and surface missing information.
Conflicting ownership
Re-check that a document belongs to the sync before updating it. Preserve native content and stop when ownership cannot be established.
False success
A process finishing does not prove the destination is correct. Use execution receipts and an independent readback of the outputs being validated.
Documented validation.
Email workflow tests
Functional and simulated-fault checks passed locally, in staging, and against installed source on September 8, 2026.
Knowledge-sync tests
A separate operations suite covered ownership, pagination, failed reads, uncertain writes, and receipts on September 9, 2026.
These are dated engineering checks, not model-accuracy or uptime measurements. Mail mutation faults were simulated. The two suites are separate.
The knowledge-sync verification independently checked one changed document at its destination. That execution also reported 833 unchanged files and zero errors. One verified document does not establish every document’s correctness.
My contribution.
I define the requirements, workflow architecture, operating rules, and acceptance criteria. I direct development with Claude Code and Codex, review the resulting behavior, and use test and execution records to decide what to change next.
Olympus uses existing models, tools, and agent runtimes. It is an independent personal implementation; no company-wide productivity gain or enterprise rollout is claimed.
Download the case study PDF ↓