Selected work · Joe Cochran

Different systems.
One working flow.

Three systems I built in previous roles. Each connects information, people, and decisions across the gaps where work usually slows down.

Fleet citationsField-service qualityMulti-currency payments

Fleet citations

Paperwork → people → payment

Different citation formats. One question before payment: who had the vehicle?

Illustrative workflow · not a client screenshotFictional sample data
  1. CITATION ANOTICE BTICKET C
  2. Vehicle V-042Date 12 May · 10:30
  3. 09:00 ━━ 10:30 ━━ 17:00V-042 → Alex (sample)
  4. Human review↓ approval requiredPayment handoff

Uncertain reading or assignment? Route to a person—not straight to payment.

Explore each step to see how the process connects.

Where the value comes fromConnect the paperwork to the responsible person and approval decision, with less repeated handling.

Read the full story

The challenge

A vehicle record alone wasn’t enough. The citation needed to be connected to the person assigned to that vehicle at the relevant time, then routed through notifications, approvals, and payment processing.

What the Engine Built

A workflow combining OCR, vehicle lookup, assignment cross-referencing, notifications, payment-system integration, and an approval gateway. The value was in connecting the whole process—not just reading a piece of paper.

The value provided by the solution

Less repeated handling between paper intake, vehicle records, and payment processing. A clearer connection between the citation, the responsible person, and the approval decision.

What the Engine could add next

Roll citation costs into per-driver and per-route reporting, so repeat patterns show up before they become a budget line. Flag vehicles or routes generating citations at an unusual rate, and extend the same approval gateway to related fleet expenses like toll violations or impound fees.

The broader pattern: connect incoming paperwork to the records, people, and decisions needed to act on it.

See how this conversation actually wentAn invented but realistic transcript—not a real client exchange, but the kind of back-and-forth that shapes a build like this.
User

We manage a fleet of about 40 vehicles, and we get 200+ parking and toll citations a month—different formats depending on which city or agency issued them. Right now someone manually reads each one, figures out which vehicle it's for, figures out who had it that day, and routes it for payment. It eats a ton of time and stuff falls through the cracks.

Engine

A few questions before I scope it: how are citations coming in today—mail, email, a portal? And how do you currently track which driver had which vehicle on a given day?

User

Mostly mail, some email PDFs. Vehicle assignments live in a shared spreadsheet—drivers check vehicles in and out.

Engine

Got it. And when a citation comes in but the assignment record is unclear—say two drivers overlap, or the log wasn't filled in—what happens today?

User

Honestly, it usually just sits until someone has time to dig into it, or it gets pinned on whoever's currently driving that truck, which isn't always fair.

Engine

Here's what I'd build: read each citation regardless of format, extract the vehicle and date/time, cross-reference your assignment records to find who had it, and route it for approval before anything reaches payment. Where the reading is uncertain or the assignment is ambiguous, it goes to a person—not straight to payment.

User

Who approves the payment?

Engine

That's one thing I need from you—who actually has authority to approve a citation payment, and does that change by amount? I'll build the approval gate around how it really works, not assume one approver.

User

Ops manager for anything under $200, me above that.

Engine

I'll build the gate at that threshold. I'll pull a few months of your real citations and assignment records so I can test extraction and matching against your actual formats before anything touches a live payment.

Continued

Engine

One more thing on the approval side—who exactly is "the ops manager" here? I'd rather route this to a small distribution list than one person's inbox, so nothing stalls if they're out sick or on vacation.

User

Good point. It's usually Maria, but Dave covers when she's out.

Engine

I'll set it up as a shared distribution list—fleet-approvals@yourcompany.com—with Maria and Dave both on it, so either can approve and you're not blocked by one person's inbox. Sound right?

User

Yeah, that works.

Engine

Now, on the intake side—where should these citations actually land? A dedicated inbox I monitor, or an existing one?

User

Let's set up a dedicated one—citations@ourcompany.com or something. Easier to forward mail there too if something comes in physically.

Engine

I'll watch that inbox specifically, so nothing else in a shared mailbox gets touched by mistake. Two things I want to handle carefully there: duplicates, and citations I can't confidently read.

On duplicates—agencies sometimes resend the same notice, or a citation gets forwarded twice by different people. I'll check incoming citations against citation number, vehicle, and date before creating a new record, and if it looks like a repeat, I'll link it to the existing one instead of starting a second approval chain.

User

Good, because that's happened before—same ticket, two people forward it, and we almost paid it twice.

Engine

That's what the duplicate check is for. And if a citation comes in and I can't reliably read it—bad scan, handwriting, a format I haven't seen before—I won't guess. It goes into a "needs a human look" queue with the original attached, and I'll flag it by email rather than silently dropping it or blocking everything else behind it.

User

That's fine, as long as it doesn't just get lost.

Engine

It won't—nothing leaves the queue without either being processed or explicitly reviewed by a person. Since we're talking about visibility—do you want a standing aging report? Something that shows citations sitting in review, how long they've been there, and which ones are getting close to a late-fee deadline?

User

Oh, actually—that would be really useful. We've eaten late fees before because something sat too long.

Engine

I'll build that in as a weekly aging summary. Who should that go to?

User

Same distribution list, plus me on Fridays.

Engine

Got it. Last piece—how do you want these actually paid once they're approved? Should I post directly to your payment system, or does someone need to review the transaction first?

User

Create the entry in QuickBooks, but don't pay it automatically—I want a person to actually look at it before money moves.

Engine

Understood—I'll create the bill in QuickBooks on approval, flagged for payment, but the actual payment step stays a manual action on your end. Nothing pays itself.

User

Perfect.

Engine

Before I build against your real data—can you forward one sample citation, whatever format is easiest, to citations@ourcompany.com so I can see exactly what I'm working with?

User

OK, sent.

Waiting…

Engine

Got it—that came through as a PDF from the city's Parking Authority. I can read the citation number, vehicle plate, and violation date and time clearly. The issuing agency's letterhead is in the top corner too—I'll use that to help identify the format going forward. One thing I want to flag: the fine amount on this one is handwritten in a stamped box, not printed—that's the kind of thing I'd rather send to a person to confirm than guess on. Does that match how these usually look?

User

Yeah, that's pretty typical actually—the amount is often stamped by hand.

Engine

Then I'll treat a handwritten or stamped fine amount as an automatic "needs review" trigger, even if everything else on the citation reads cleanly. Better to have a person confirm a dollar figure than have me guess at one.

Have a similar challenge? Request a consultation →

Field-service quality

Spot the mismatch before submission

Why wait for downstream QC if similar, approved work is already available?

Illustrative workflow · not a client screenshotFictional sample data
  1. “Replaced damaged cable”Draft code INSTALL
  2. “Repaired cable fault”“Replaced damaged line”Prior code: REPAIR
  3. INSTALLCompare withREPAIR
  4. Technician checks context↓ chooses the right codeSubmit job report

A new or unusual job may have no good historical match. Keep a route for human judgment.

Explore each step to see how the process connects.

Where the value comes fromPut relevant prior work in front of technicians while they can still address inconsistencies.

Read the full story

The challenge

How do you use that history to identify potential coding problems before submission, instead of relying entirely on downstream quality checks?

What the Engine Built

A pre-submission fuzzy-matching check comparing a technician’s description against historical descriptions and approved timecode combinations. It used similarity to surface likely matches while the technician was still completing the report.

The value provided by the solution

Put relevant prior work in front of the technician at the point of entry. Create an opportunity to catch mismatches before a separate QC cycle and its follow-up conversations.

What the Engine could add next

Surface technicians whose coding consistently drifts from their peer group, turning it into a targeted training conversation instead of a recurring QC finding. Feed coding-accuracy trends back into scheduling, so the right technician gets matched to unfamiliar job types.

The broader pattern: use accepted prior work to help people make better decisions before mistakes travel downstream.

See how this conversation actually wentAn invented but realistic transcript—not a real client exchange, but the kind of back-and-forth that shapes a build like this.
User

Our field techs log their own time codes after each job—stuff like INSTALL, REPAIR, MAINTENANCE. Problem is they're often wrong, and we don't catch it until QC reviews it days later. By then it's a whole back-and-forth to figure out what actually happened.

Engine

How many jobs a week, and what does the current QC process actually look like—spot check, full review, something else?

User

Around 150 jobs a week across the team. QC pulls a sample, maybe 20%, and flags anything that looks off based on the write-up.

Engine

So most miscoded jobs never get caught at all.

User

Right. And the ones that do get caught, it's usually days later, and then we're chasing the tech down to ask what actually happened.

Engine

I'd move this to the point of entry instead—compare what the tech writes against your history of already-approved jobs, and surface likely matches before they submit. Two things I need: do you have a searchable archive of past job descriptions and their final approved codes, and how far back does it go?

User

We've got about three years in the system. Not perfectly clean, but it's there.

Engine

That's enough to work with. One thing to be clear on up front: a similar-sounding description isn't proof the code is right—it's a prompt for the tech to double-check. I don't want this making the call for them.

User

Agreed, we don't want it auto-correcting anything. Just flag it.

Engine

Then here's what I'd build: the tech writes their description and picks a code like normal. Before they submit, I compare it against similar historical jobs. If the code they picked doesn't match what similar jobs used, I show them the comparison and let them decide—submit as-is or change it.

Continued

User

What happens with a job that's genuinely new—nothing like it in the history?

Engine

No flag at all. If nothing similar exists, I stay out of it rather than force a comparison that isn't useful. That's actually the harder failure mode to get wrong—flagging things that don't deserve it trains people to ignore the flags entirely.

User

Makes sense.

Engine

I'll need a couple weeks of your real job descriptions and codes to tune the matching before this touches anything live. Once it's running, I want to test it against a live submission—can you have a tech file a job through the normal app once I've got this wired in, using a description close to something you know should trigger a flag?

User

Yeah, I'll have Marcus run one tomorrow.

Waiting…

Engine

Marcus's job came through—description was "replaced faulty line," he coded it MAINTENANCE. Closest historical matches used REPAIR. The flag triggered correctly and showed him both prior examples before he submitted.

User

Did he change it?

Engine

He kept MAINTENANCE. I don't have visibility into why—that's the point, it's his call, not mine. Worth asking him directly whether that was intentional or the flag didn't register.

User

I'll check with him.

Engine

One more thing worth deciding now: how often should I refresh the matching against new approved jobs—weekly, monthly?

User

Monthly's fine, we don't get that much new job-type variety.

Engine

I'll set it to pull a fresh set of approved jobs on the first of each month.

Have a similar challenge? Request a consultation →

Multi-currency payments

See both sides of the payment flow

Customer invoices and vendor obligations span up to five currencies. Fees belong in the decision.

Illustrative workflow · not a client screenshotFictional sample data
  1. USDEURGBPCADJPY
  2. EURUSDCADJPYGBP
  3. Customer side ↔ Vendor sideBank transaction costsOperational constraints
  4. Chosen approach↓ processingPayment workflow

Illustrative currencies only. No exchange rates, savings estimates, or proprietary algorithm are shown.

Explore each step to see how the process connects.

Where the value comes fromMake transaction costs part of a repeatable processing decision across customer and vendor flows.

Read the full story

The challenge

Customer-side and vendor-side transactions have to be considered together. A convenient choice for one payment may not be the most cost-effective choice across the operation.

What the Engine Built

Software supporting multi-currency transaction processing on both sides of that flow, with minimizing bank transaction fees as the objective. This was a payment-operations problem—not a reason to add AI for its own sake.

The value provided by the solution

Make transaction costs part of the processing decision instead of treating each leg in isolation. The objective is a more cost-conscious workflow, not a claim of guaranteed savings.

What the Engine could add next

Once real fee and balance data is flowing, forecast upcoming FX exposure and flag when pre-purchasing a currency ahead of a rate move is worth considering. Reconcile the audit trail against actual bank statements each month, automatically.

The broader pattern: turn a complicated operational decision into an explicit, repeatable process.

See how this conversation actually wentAn invented but realistic transcript—not a real client exchange, but the kind of back-and-forth that shapes a build like this.
User

We do business in five currencies—USD, EUR, GBP, CAD, JPY—and we're constantly moving money both ways: customers pay us in their currency, and we pay vendors in theirs. Right now everyone just picks whatever's convenient for a given payment, and I have a feeling we're bleeding money on fees without really seeing it.

Engine

Walk me through a typical week—how many payments are we talking about, and who currently decides how each one gets sent?

User

Maybe 30-40 transactions a week combined. Whoever's processing AP or AR that day just uses whatever method is already set up for that vendor or customer—usually a bank wire or whatever the accounting system defaults to.

Engine

So there's no one actually comparing the cost of routing options against each other?

User

No, not really. We just want it to go out or come in without errors. Nobody's tracking what we're losing to fees or bad exchange spreads.

Engine

That's the gap I'd build against—not adding AI for the sake of it, but treating this as a payment-operations problem: look at customer-side and vendor-side together, and make transaction cost part of the actual decision instead of treating each leg in isolation. Before I scope the mechanics, I need a few specifics: where do the actual payment deadlines and available balances live today—a bank portal, your accounting system, something else?

User

Mostly our accounting system tracks due dates. Balances we check directly in each bank's portal since we bank with a couple different ones for different currencies.

Engine

Got it. And do you have visibility into the actual fee schedules—like what your bank charges for a wire in each currency—or does that live in someone's head, or an old email from your banker?

User

Ha, pretty much the second one. We know roughly, but nobody's got it written down anywhere current.

Engine

That's the kind of thing I'd rather nail down with you before building anything than guess at. I'm not going to invent exchange rates or promise a specific savings number; the actual approach depends entirely on your real fee schedules, balances, and deadlines. What I can build is the structure: pull customer invoices and vendor obligations together across all five currencies, layer in the real fee data once we have it, and turn "which way do we send this payment" into a repeatable decision instead of a habit.

Continued

Engine

Let's get into the mechanics. For pulling in the actual transaction data—do you want me connecting directly to your accounting system's API, or is there a simpler export/import you'd rather start with?

User

We're on QuickBooks for the US side, and a separate system for the international entities. Can you handle both?

Engine

I can—I'll need read access to each, and I'll want to understand where they might disagree with each other, since multi-entity setups sometimes have the same transaction recorded slightly differently in each system. When that happens, which one is the source of truth?

User

QuickBooks, honestly. The international system is more for local compliance.

Engine

That settles a lot of edge cases upfront. I'll need read access to both systems set up before I go further—can you get that provisioned?

User

I just added the read-only credentials for QuickBooks. Can you see it?

Waiting…

Engine

Confirmed—I can see the chart of accounts and the last 90 days of AP/AR activity. I count 214 open vendor bills and 87 open customer invoices across the five currencies. Access is working; I still need the international system's credentials before I can compare the two.

User

I'll get Priya to send those over this week.

Engine

Now, on the bank fee side—since that data isn't written down anywhere current, I'd rather not hard-code numbers that'll drift out of date. What if I build a simple place for you to enter and update fee schedules per bank and currency, and flag it for review on a set schedule—say quarterly—so it doesn't quietly go stale?

User

That works. Who should get that reminder?

Engine

Whoever owns the banking relationships—is that you, or does someone on your finance team handle that day to day?

User

That's Priya, our controller.

Engine

I'll route the quarterly fee-schedule review to her. Now, on actually executing a payment once the system recommends a routing—does it just go out automatically, or do you want a person confirming before money moves?

User

No, definitely a person. This is real money moving internationally—I want someone looking at the recommendation and approving it, not a system just sending wires on its own.

Engine

That's the model I'd push for regardless of what you'd said—I'll surface the recommended routing with the reasoning behind it, why this path costs less than the alternative, and a person approves before anything executes. Nothing sends itself.

User

What happens if two payments are due the same day and the balances don't support the "cheapest" routing for both?

Engine

That's a real constraint, not an edge case. When there's a conflict like that, I won't silently pick one; I'll flag both options with the tradeoff spelled out—for example, "cheapest routing for Payment A conflicts with Payment B's deadline; here are the two ways to sequence this"—and let a person make that call, since it might depend on things I don't have visibility into, like which vendor is more forgiving about a day's delay.

User

That makes sense. One more thing—can you keep a record of what we actually decided over time? I'd like to eventually see if this is actually saving us anything.

Engine

Yes—I'll log every recommendation, what was actually approved, and why, so you've got a real audit trail. Once we've got a few months of that, we can look at it together and see what it's actually telling us, instead of guessing.

Have a similar challenge? Request a consultation →