Eric Solak

The hard part of automating their job costing wasn’t the code. It was learning a process that lived in the bookkeeper’s head, from someone who did it by feel and had never had to explain it.

How I ran discovery

• 2–3 calls with the bookkeeper to learn the company’s workflow and what they needed from the automation.
• Walking through finished job folders together, bill by bill.
• Each folder held invoices, quotes, and freight, customs, site work and surfacing bills, as PDFs, scans and Word files.

What I asked

• “Walk me through this job. For each bill, what do you calculate, and where?”
• “Why is the tax on this bill treated differently from that one?”
• “Which exchange rate do you use, and where does it come from?”
• “What do you do when a bill hasn’t arrived yet? How would you know one is missing?”

What I learned

Walking through real jobs turned up four rules:

• Sales tax differs by vendor. Some Canadian sales tax (HST) can be claimed back, one carrier charges none, and customs bills mix fees, duty and import tax that are each treated differently.
• The exchange rate is theirs, not the market’s. It must match what their bank charged, so an automatic rate lookup would have been wrong.
• Every number must trace to a document in that job’s folder. Nothing is estimated.
• The folder is never complete. A quote often stands in for an invoice that hasn’t arrived, and a missing bill is only noticed when another document mentions it.

What looked like case-by-case judgment was four firm rules that nobody had written down.

How it shaped the build

Each finding became a design rule:

• Every number traces to a document, so the AI agent only reads the bills and plain Python does every calculation. The bookkeeper can audit any figure.
• Tax differs by vendor, so the vendor rules live in one file. A policy change is a single edit.
• The exchange rate is theirs, so the bookkeeper types it in. It is never looked up.
• The folder is never complete, so quotes standing in for invoices, bills that other documents mention, and any rule I was unsure of are flagged for review instead of guessed.

Impact

Hours of typing per job became minutes plus a short review list. Together with my other automations for them, the client saves 26+ hours a month. Once running, the automation also flagged a $3,000+ government rebate that had gone unclaimed for more than a year.

“We used to dread doing this because it would take hours out of our day. Now we are so far ahead on our work.”The client

Next: scaling it up

Version one proved the rules with the bookkeeper in the loop. With that trust in place, the next phase is:

• Pulling bills straight from the shared inbox, so each job is costed as its bills arrive instead of waiting for a folder to be filled.
• Logging every flag the bookkeeper overrides, so the written rules tighten from real corrections as volume grows.
• Reusing the pipeline for other distributors. The rules already live in one file, so a new client means writing down their rules, not new code.

Next write-upVoice tasks for dairy farms, tested with the people who use them

Pocket AI turns one spoken sentence into a task for dairy farm workers. I led AI development on our five-person capstone team and wrote our usability report. The challenge was building for people who work with their hands full, in loud barns with weak signal.

Setting requirements with the client

The CATTLElytics team set scope and priorities with us in regular Teams meetings. That gave us four requirements:

• Voice first. Operators have their hands full, so a task should take one spoken sentence, not a form.
• Built for the barn. Recognition had to reach 95% over barn noise.
• Works without signal. Farms have weak connectivity, so tasks had to save without a connection and sync later.
• A focused first release. Multi-language, wearables and advanced NLP moved to stretch goals so the core loop could ship.

What we built

OpenAI Whisper transcribes speech and a language model pipeline pulls the description, deadline, assignee and priority out of a single sentence. A Flask backend handles authentication, task lifecycle and sync with CATTLElytics’ proprietary storage. The app is offline-first: tasks save to local SQLite, sync in the background, and show when they are still waiting to sync. Tasks are assigned to people by name and can repeat on a schedule. Every task goes through a confirmation step before it saves, and anything can be edited by touch.

How I validated it

• The users: five dairy farm operators, two comfortable with voice assistants and three not.
• The tasks: operators created real tasks, such as “check Cow 42 for lameness tomorrow morning,” three times each, unaided, with recorded barn noise playing.
• The debrief: each operator talked through what worked and what they would change.

Results

Recognition held at 98% under simulated barn noise, beating the 95% target. Speech to confirmation took 5.8 seconds on average, and operators rated their overall satisfaction 4.4 out of 5. Saying the description, deadline, assignee and priority in one sentence was what all five liked most. The app was handed off to CATTLElytics for integration into their platform, and the project was highlighted on official McMaster channels.

What testing surfaced

• Tapping to talk got in the way. Three of five had their hands full with animals or equipment when they needed to start a task.
• Unsynced tasks looked lost. Two of five expected a new task to appear right away, before it had synced.
• Everyone wanted names, not IDs. All five asked to assign tasks to people by name.

Next: what I handed off

My usability report closed with a roadmap for the integration:

• A wake word, for fully hands-free use.
• Field testing in live barns with more operators.

After the capstone, I mocked up a redesign that answers each finding. It is a concept, not the shipped app: see the Pocket AI redesign concept

Next write-upAn LLM pipeline a bookkeeper can audit