AI operations · Commerce & marketplaces

Taking AI audits for 2026

AI that earns
its keep.

I'm Jason. I put AI into production for commerce and marketplace teams — support triage, catalog enrichment, and internal operations. Narrow scope, real evals, human fallback.

Selected work

AI intake at Inspectrik. Commerce operations at Thinkware US and Canada.

Most AI projects die between the demo and the workflow. I fix that: one narrow use case, real data, explicit evals, a human fallback, and a cost ceiling. Then we scale what survives.

AI operations

Three workflows. Measurable returns.

Narrow workflows, real data, explicit evals. Start with one track; expand only when the numbers hold.

01

Support triage

Deflect the repeatable. Escalate the rest. Classify tickets, retrieve the right answer, draft a response, and route edge cases to a human with full context.

  • Eval set built from real tickets
  • Retrieval over help centre and orders
  • Confidence threshold and escalation
  • Cost per resolved ticket
See the service
02

Catalog enrichment

Turn product data into structured data. Extract attributes, normalise taxonomy, write descriptions, and localise catalogues without losing SKU truth.

  • Attribute extraction and QA sampling
  • Taxonomy and dedupe rules
  • US and Canada localisation
  • Bulk pipeline with review queue
See the service
03

Internal operations

Give the team a copilot with guardrails. Document Q&A, vendor and ops assistants, and approval workflows with a human on every irreversible action.

  • Retrieval over SOPs and documents
  • Tool access with approvals
  • Audit log and retention policy
  • Model routing and budgets
See the service

How I work

Built for production.
Engineered to ship.

Four commitments held on every AI engagement, from the first audit to the model change that ships on a Friday.

Workflow

Automate a job, not a chat.

A queue, an owner, and a measurable outcome. No chat box looking for a use case.

Evals

Prove it before launch.

Labelled cases, failure budgets, and regression checks on every model change.

Data

Know what leaves the building.

Retention, redaction, and vendor boundaries agreed before the first prompt.

Cost

Set the ceiling first.

Token budgets, caching, routing, and a human fallback when the meter runs.

Selected work

Production proof.
Not a demo.

One AI system live, two commerce operations feeding it. Each relationship direct, ongoing, and accountable to a real workflow.

01 · Commerce dataOngoing

Thinkware Store — US

Commerce data foundation for AI: catalog attributes, support macros, and storefront integrations. The plumbing an assistant needs before it can be trusted.

CatalogStructured data
SupportMacro library
LiveStorefront
Read the case study
02 · Multi-marketOngoing

Thinkware Store — Canada

Second market on the same data contract: regional pricing, catalog parity, and checkout. Consistency across storefronts is the prerequisite for automation.

2Markets
ParityCatalog
OngoingOperation
Read the case study
03 · AI intakeOngoing

Inspectrik

AI-assisted intake across chat, video, and in-person inspections. Guided triage routes requests, structures summaries, and keeps a human on every recommendation.

AI triageIntake
3 modesChat · video · visit
HumanReview
Read the case study

Tracks

Scoped

3tracks

Support, catalog, operations

Three production workflows. Start with one audit; expand when the numbers hold.

In production

1system

AI intake at Inspectrik

Guided triage across chat, video, and in-person inspection, with human review.

Foundation

US / CA

2platforms

Commerce data in production

Thinkware US and Canada: catalog, support, and storefront operations.

Human in the loop

Every irreversible action has an owner.

Clients

Long-term relationships. Not one-off projects.

Three active clients across commerce and property tech. Each relationship is direct with me, and each one runs in production.

Active
03
Markets
US / Canada
Model
Direct retainer
Platforms in play
WORDPRESSWOOCOMMERCECLOUDFLAREMARKETPLACE OPSAI INTAKE

Notes

Latest notes.

Occasional writing on evals, data boundaries, and the unglamorous middle of putting AI into production.

All notes