Cogs Consulting

Case studies

Examples of systems I've built, and what changed for the businesses running them.

01

Anvil Coffee Collective

Coffee · Multi-site · Ongoing · Salmon Arm, BC

They ran on a point of sale, a scheduling tool and an accounting package. Each one worked fine on its own, but none of them could see the others, so anything that needed two at once got done by hand or didn't get done. Those systems are now connected — none of them replaced.

  • A custom MCP server connects their own systems to Claude. They ask a question in plain English and get an answer from their real numbers, without opening any of the tools underneath.
  • The weekly prep numbers used to take about an hour. They take about thirty seconds now, worked out straight from sales.
  • Tip distribution pulls tips from the point of sale and hours from the scheduling tool, then applies the shop's own rules. It runs the same way every week no matter who runs it.
  • Meeting transcripts become tasks in the tool the team already works from, so what gets decided doesn't rely on someone remembering it.
  • Sales, labour and costs sit in one place, so a question that needs all three finally has an answer.

The result is a view of their own business that most small businesses never get — no full-time analyst, no enterprise reporting suite. Decisions that used to come down to a hunch now start with the numbers.

02

Valley Tool Repair

Tool repair shop · Sole operator · 2026 · Summerland, BC

Inventory lived in a spreadsheet that only stayed accurate if it was updated after every repair. Ordering meant walking into the shop to recount parts, then opening supplier websites one at a time to find the best price. It's now one system covering inventory, invoicing, and purchasing.

  • Finalizing an invoice deducts the parts used, and reverting one puts them back. Nobody counts stock by hand.
  • Every supplier's price for a part sits in one row, so the cheapest option is right there on the screen he's already on when ordering.
  • Freight is spread across the parts on an order when it arrives, so what a part costs on paper is what was actually paid for it.
  • The reorder list fills itself from stock minimums and turns into a purchase order in one step.
  • Customers, invoices, supplier prices and parts diagrams live in one place instead of across binders, folders and spreadsheets.
  • Years of old invoices were imported at setup, so none of the history had to be retyped.
  • There's no AI in the system he uses. It's ordinary software doing the same arithmetic every time.
"I had a guy just call asking for a part for a certain tool, and I was able to check the diagram, locate the part in inventory and bill him — all right on the phone."— Owner, Valley Tool Repair

Also built

Additional work, described in technical terms rather than client-specific detail to protect client privacy.

Video tagging and metadata

Vision analysis runs over full-length footage for a media company and returns structured output constrained to an approved tag vocabulary, so the model can choose only from values the company actually uses. Every result is validated in code against the company's written production standards before it is written back to the asset management system. Violations are flagged for human review, and hand-edited fields are preserved on write.

Hosted document search (RAG)

Large document collections are ingested at scale, chunked, and embedded as vectors in a Postgres database, so questions are answered through semantic retrieval rather than keyword matching. The system finds the passage that answers the question even when the wording doesn't match, then returns answers grounded in the source text with citations instead of a list of files to open. Built to be hosted and maintained on your behalf: you hand over the documents, and your team gets a search interface over material that was previously unsearchable.

Editorial and analytics automation

For an online news publication. A custom MCP server connects three separate data platforms — search analytics, site analytics, and keyword and ranking data — into a single interface, so a question that would normally mean three dashboards and a spreadsheet is one query. A companion tool pulls candidate stories from a wire service, runs each through an editorial checking pass, and formats the result for human review. Validation is enforced in code around every generated step, and the whole thing runs as a shared TypeScript monorepo, so a capability added once is available everywhere rather than rebuilt per app.

Where to start

Most engagements start with a short conversation about what you're doing by hand, then a working prototype you can react to. Every build ships with documentation so you can carry on without me.

Tell me what the annoying part of your week is.