Rohan Jasani

Rates, paperwork and a spreadsheet nobody wanted to give up

Level Up is a small mortgage brokerage. When I arrived for a summer internship from UBC, nobody handed me a spec. So I asked the owner and the operations lead one question: where is the week actually going?

Then I followed a deal through the office to find out. A lot of skilled time was going into chasing rates, sorting paperwork, and working out which documents were still missing. None of that was the job. All of it still had to be right.

What I built in response is Level Up’s internal AI operations platform. The team calls it Lum Brain. It connects the tools the brokerage already relied on, including Gmail, Zoho CRM, Finmo, Google Drive and Slack, and takes over the connective work between them.

A frame from the tour: a sample client’s documents arrive, and the platform names and files each one. Watch the 3-minute tour
The quickest way to understand the platform is to watch it work. The tour runs three minutes, and everything on screen is sample data.I made the tour with Margin, a tool I built for turning a script into a product walkthrough.

Rates that keep themselves current

A broker’s advice is only as good as the rates behind it. The team’s usual rate lookup tool didn’t cover every lender the brokerage works with, or the detail behind each rate, so the gaps were filled by hand in a spreadsheet. Rates change constantly, the spreadsheet went stale, and the tool offered no way to pull its data out.

The frustrating part was that every rate we needed was already arriving by email, including from the lenders the tool left out. Reading them was the problem. A lender’s rate sheet is a dense grid, often sent as an image or a PDF. Terms run down the side. Insured, insurable and uninsurable run across the top. The numbers shift again with how much the client is putting down, and the fine print decides which of them apply.

A rate sheet, much simplified

Alder Trust Broker rate sheet, effective November 3

Term Insured Insurable Uninsurable
1-year fixed 5.24 5.34 5.49
3-year fixed 4.44 4.54 4.69
5-year fixed 4.24 4.34 4.49
5-year variable P-0.85 P-0.75 P-0.60

Promotion ends November 30. Owner-occupied purchases only. Add 0.10 for amortizations over 25 years.

The lender and the numbers are invented. A real sheet carries dozens of rates, and every one has to land in the right cell, for the right lender, with its conditions attached.

Keeping the spreadsheet current meant reading each rate off a grid like that and typing it into the matching cell, then doing it again the next time a lender moved a number. It is careful, spatial work, and nobody at a brokerage has spare hours for it.

So I built a system that reads each lender’s emails and attachments and keeps a living rate card for that lender: the rates, the promotions and when they end, and the policies that decide who qualifies. When a new sheet arrives, the rates it replaces are retired. It follows 66 lenders and has revised their cards more than 1,200 times.Counted from the production database on October 5, 2026: 66 lender rate cards and 1,267 revisions. I ran totals only. No client details were read.

One rate email, start to finish
  1. The system spots a rate email in the morning’s mail.
  2. The system reads the sheet, including the image or PDF attached to it.
  3. The system checks every new rate. One that looks wrong never reaches the card.
  4. The system updates that lender’s card, retires the old rates, and posts what changed in Slack.
  5. A person reviews anything the checks held back.

When the team builds a proposal, the comparison starts from those cards. So the suggestion is specific: the sharpest rate available, whether the brokerage can place a deal with that lender, and what the client needs to qualify.

The paperwork

Client documents arrive by email, upload and share link, usually with names like IMG_2207. The platform reads each one, works out what it is and whose it is, gives it a proper name, and puts it in the right client’s folder. Anything it is unsure about waits for a person. In its first three months it filed more than 3,900 documents.Counted on October 5, 2026: 3,925 documents since July 10. On filing alone, that gave back an estimated 8 to 12 staff hours a week.This one is an estimate, and it covers filing only. It assumes 1.5 to 2.5 minutes to download, rename, find the folder and upload each document. Across filing, CRM notes and rate upkeep my estimate is 250 to 400 hours over about four months, and that still leaves most of the automations out.

Below is a small working version of it, with invented names. It files what it is sure about and asks a person about the rest. Here, the person is you.

Just arrived

Five documents are waiting to come in.

Waiting for a person

Nothing is waiting.

Client folders

  • Daniel Okafor
    • Maya Lindqvist
      • Magnus Lindqvist

        Five documents are on their way in.

        The version nobody used

        My first deal view was called Client 360. The idea was modest: a faster way to see what the CRM already knew about a client. I gave it a clean, minimal layout built from cards, and I thought it was better organized than what the team had.

        Adoption trailed anyway. The person who ran every deal kept using the spreadsheet she had relied on for more than three years, even though it was more work. A new layout was friction, and a three-year habit beats a tidy screen.

        Two ways to show the same deal

        What I built first. Tidy cards. Easier to read, and new.

        What the team trusted. Rows in a sheet. More work, and theirs.

        So I stopped trying to replace the spreadsheet and built a better one. The planner looks and behaves like the sheet the team already knew, with ideas borrowed from Notion and Airtable and none of the setup.

        What the planner can do that a spreadsheet can’t

        A mortgage approval comes with conditions, and most conditions are documents: pay stubs, a T4, a void cheque. In the planner, every condition is a row. Over the weeks a deal takes, the platform collects the client’s documents, files them, and links each one to the condition it answers.

        Open a row and the document is there beside the checklist, behind the same sign-in as the rest of the platform, so nobody hunts through Drive or forwards an attachment. At a glance the team can see which conditions have their documents, which are waiting on the client, and which are with the lender. Each row also shows the dated reason it sits where it does: the email that went out, the document that came in, what the lender said.

        One lender condition, start to finish
        1. The lender asks for two recent pay stubs.
        2. The system adds a row to the checklist, in the lender’s own words.
        3. The client emails the pay stubs a few days later.
        4. The system files them, links them to the row, and hands the row to the team.
        5. The team sends them on to the lender.
        6. The lender accepts them. Only now does the row close.

        A filed document moves a condition forward. Only the lender’s word closes it.

        People still decide

        Changes the AI wants to make to client records are staged first, with an audit trail. The low-risk kinds moved to automatic approval only after the team had watched them be right for a while. No email goes out without a person sending it. And SINs and bank, card and account numbers are stripped from emails and notes before they reach an AI model.Every AI task can also be switched to run on Gemini in Montréal with a settings change. It is a ready Canada-only path if the rules on data ever get stricter.

        One change to a client record, start to finish
        1. The system drafts a note for the client’s CRM record.
        2. The system stages the draft. Nothing has changed yet.
        3. A person reads it and approves it.
        4. The system writes the note, marks it as AI-written, and logs all four steps.

        Orange is the system and teal is a person, in every figure on this page. Outgoing email never skips the teal step.

        Still running

        Today the platform runs 99 scheduled automations, which have run more than 145,000 times at a 99.1% success rate.Also counted on October 5, 2026: 145,394 runs, 1,300 failures. Besides rates, filing and the planner, it keeps client records current, turns meeting recordings into notes, and drafts the weekly rate promotion for someone to approve.Underneath it are more than 5,000 automated tests. I build with AI coding tools, and the tests are how I know what they wrote is right.

        I wrote a plain-language guide so the team can run and audit all of it themselves, and built troubleshooting help into Ask Lum, the assistant they use in Slack. I still support them part-time as new needs come up.

        Level Up’s public page for the platform. A sample lender rate sheet sits beside the answer the assistant drew from it.
        I also designed and built the public page that explains the platform to Level Up’s clients. It runs on sample data. Visit levelupmortgages.ca.

        Email me about this Back to the home page