Selected work

Production automation, not demos.

Each of these ran against a real system of record, with real consequences for getting it wrong. Client names and identifying details are withheld; the mechanisms are described exactly as built.

Identity lifecycle automation

1,000+ employee company, internal IT

Boss Level

Situation: Group membership lived in the corporate directory, but the chat platform's user groups were maintained by hand. Every reorganization meant someone re-checking dozens of groups, and access that should have lapsed on a Friday often survived for weeks.
What I built: A scheduled sync between the directory and the chat platform's user groups, comparing desired state against actual and reconciling the difference. Critically, it runs dry-run first by default: every proposed add and remove is reported before anything is applied, so a bad directory change cannot cascade into mass access removal.
Outcome: Group membership became a consequence of the directory rather than a separate manual chore, and offboarding stopped depending on someone remembering.

n8n
Microsoft Entra ID
Slack API
Scheduled reconciliation

Recruiting pipeline with a human gate

High-volume hiring, talent operations

Co-op Mode

Situation: Screening a large applicant pool against role criteria consumed hours of recruiter time per requisition, and the criteria were applied inconsistently depending on who was reviewing and when.
What I built: An automation that pulls applications from the applicant tracking system, evaluates each against the role's stated criteria, and produces a structured recommendation with reasoning. Every candidate-affecting action sits behind an explicit human approval step — nothing is advanced or rejected without a person clicking approve.
Outcome: Consistent first-pass screening with a full audit trail, and recruiter attention spent on judgment calls rather than on reading every résumé from scratch.

n8n
Greenhouse ATS
LLM evaluation
Human-in-the-loop approval

Support-to-fulfillment pipeline

Operations, inbound request handling

Co-op Mode

Situation: Inbound requests arrived as support tickets containing structured data that a person then re-typed into a downstream inventory system. Slow, and every transcription was a chance to introduce an error.
What I built: Extraction of the relevant fields from each ticket, validation against the downstream system's requirements, and automated upload — with anything ambiguous routed to a person instead of guessed at.
Outcome: Manual re-entry eliminated for the well-formed majority of requests, with exceptions surfaced explicitly rather than silently mishandled.

n8n
Freshdesk
Inventory system API
Validation and exception routing

The common thread

All three touch a system of record

Directory groups, candidate records, inventory. These are the integrations most automation shops avoid, because the failure modes are expensive and the APIs are unglamorous. They're also where the hours actually are.

Tell me about yours