Sheffler & Company: from a Wix page to running their proposals

The short version

Sheffler & Company has been surveying and engineering land in western Pennsylvania for more than fifty years. In 2024 they hired me to build a website.

Today the proposal process that runs their business runs on a platform I built for them, and the website is the least interesting thing I have done for this client.

This is the story I point to when someone asks what happens after the website.

The website

Bud Sheffler runs a firm his family name has been on for half a century. Licensed surveyors and professional engineers, work across Pennsylvania, Ohio and West Virginia, the kind of company that gets hired because someone’s father hired them.

Their website was a Wix page. Not broken, just small. It described a company a fraction of their size and said nothing about why fifty years of continuous operation in this field is unusual and hard.

What we built. A real marketing site: the service lines separated properly, surveying and civil engineering each given room, the project history treated as evidence rather than a list. The nicest lead form I had built at that point, with conditional logic that routed an inbound request to the right discipline based on what was actually being asked for. That mattered more than it sounds, because “I need a survey” and “I need a stormwater plan” go to different people and used to arrive in the same inbox.

Then they bought a company mid-build. Sperdute Land Surveying was acquired while the site was in progress, which meant a brand with its own history, its own clients and its own domain had to land inside a site that was not designed for it. We folded their presence into the new structure and built a bridge landing page on the acquired domain, so anyone arriving from the old brand understood what had happened rather than hitting a dead end.

The site launched in under four months and they went onto a care plan. That could have been the end of it. For most web projects, it is.

The audit they said no to

In January 2026 I came back with something bigger. I had been watching how they worked for a year and a half by then, and I could see that the interesting problem was not their website.

So I built an AI systems audit and pitched it.

They passed.

I am including this because it is the most useful part of the story. The diagnosis was right. The packaging was wrong. I was selling a category to a firm that does not buy categories, asking a fifty-year-old engineering practice to fund a study of itself. Bud goes off what the people around him say, and nobody in that room could point at a specific thing it would fix on a Tuesday.

Being right and early is still being wrong. What that no bought me was five more months of paying attention, and the next time I came back it was not with a framework. It was with the exact bottleneck, priced, with a working prototype attached.

The portal

Here is what I found once I stopped pitching and started asking.

Around 700 inquiries a year, tracked in one Excel file. Per-year tabs going back to 2022, filled in by hand by office staff. Proposal number, date received, how it came in, who took it, scope, quoted price. One file. And in the words of someone on their own team, in writing, a meaningful chunk of it lived on a whiteboard in the office.

Every proposal started with a manual research sweep. Before quoting a job, someone had to work out what the site actually demanded: flood zones, slopes, soils, watershed, road ownership, parcel boundaries, mine subsidence, wetlands, the relevant zoning ordinance. That meant visiting thirty-plus separate county and state data sources, per property, by hand. Then roughly half an hour of formatting in Word, with the copy-paste errors that come free with that.

And the risk lives in the research, not the document. This is the reframe that won the room, and it is the whole reason the project exists. Everyone assumes the expensive part of a proposal is writing it. It isn’t. The expensive part is figuring out the scope, because scope drives price, and if you get the scope wrong you either lose the job or eat the difference. Document generation is the cheap end of the problem. Due diligence is where the money leaks.

I pitched it knowledge-first and built it generator-first, in the room, in person, to Bud with Eric and Dana there. They said yes on the spot.

What V1 actually is

The Business Knowledge Base. Their institutional knowledge, extracted from three decades of proposals and the heads of the people who write them, turned into a structured asset the company owns. In its current state that is 131 scope lines across 15 services, 47 regulatory sources, 11 site triggers and 7 fee rules. This is the piece nobody asks for and everything else depends on. It is also the piece that makes them harder to replace, which is the point.

The Due-Diligence Assistant. An address goes in. Out comes the set of site signals that matter, each one carrying a clickable citation back to the live source it came from, plus pointers into the specific ordinance sections a human needs to confirm. It re-runs itself when a report goes stale.

That constraint came from Katrina, who has been doing this work for about twenty-five years and was the most valuable skeptic on the project. Her rule, non-negotiable, was that every AI output must cite a clickable source, must pull from live rather than cached data, and must keep a human in the loop. She was completely right, and designing to her standard is why the thing is trusted enough to be used.

The Proposal Generator. Approved scope becomes the finished document, on their letterhead, exported to Word and PDF.

Lead intake and the pipeline. The website form now feeds a real leads inbox instead of an inbox, without disturbing the info@ address they still rely on.

44 days

Won on June 24. All five V1 items delivered by August 7. I had quoted roughly a month, targeting under two.

296 commits. 45 database migrations. First commit the same day Bud said yes.

I want to be plain about why that number matters, because I have spent years being too modest about it. A custom multi-user platform with an AI research engine, a document pipeline and a permissions model was, not long ago, a six-figure build on a six-month timeline. It is not that anymore, for me. That is not a claim that the work got easier. It is that I have spent two years building a way of working that compresses it, and the judgment about what to build has become the expensive part rather than the typing.

What they got that they did not buy

Several things on the agreed out-of-scope list shipped anyway, because once you are inside a business the shape of the real problem keeps revealing itself:

• A lightweight CRM. Lead status board, owner and assignees, a full audit timeline per lead, editable records, and an add-never-delete model with trash and restore. This was listed as an entire separate module.
• Engineering and surveying split properly, with their own templates and their own sections on the letter.
• Save and resume, revision history, and an edit lock, so two people cannot silently overwrite each other. Ten-odd users with real roles and permissions, including a due-diligence-only role they asked for late.

And these were never on anyone’s list:

• A live document editor that matches the export page for page. The proposal is edited in a real document that matches the exported Word file and PDF exactly. That one item ate days on its own, and it is the reason the output looks like theirs rather than like software. For a firm whose letterhead has meant something for fifty years, that is not a detail.
• ALTA Table A, all twenty numbered items with working checkboxes, taken from their own form.
• A feedback widget the whole beta group actually uses, taking text, screenshots and voice memos.
• Lead attachments, AI-drafted project understanding, and create-a-lead directly from a due-diligence lookup.
• An origin-snapshot experiment. Every generated proposal is captured at the moment it is produced, so we can measure precisely what humans change afterwards and make the next one closer. Nobody asked for it. It is how the system gets better on its own.

How it is being rolled out

Not as a switch. In July we opened the beta to a core group only, the four people who actually do this work, running in parallel with the existing info@ and Excel process, with manual verification and a feedback tool wired in from day one. Their slow season is the right time to change how a business quotes work, and running both processes side by side means nobody’s livelihood depends on my software being right on the first pass.

That is deliberate. The fastest way to kill an internal tool is to force it on people before it has earned their trust.

Where it is going

Phase one is delivered. The open half is engineering, where the if-then rules that decide scope are genuinely harder than surveying and are being documented now with the people who hold them in their heads.

Phase two is everything after the proposal is composed: deposits, payments, and closing the loop into their accounting. Phase three goes further into project management and billing. There is a separate tool scoped for land-planning due diligence on a different kind of client entirely.

Why this one matters to me

I built these people a website in 2024. In 2026 I am rebuilding how they quote work, with access to their pricing logic, their institutional knowledge and their staff’s daily workflow.

Nobody gets handed that on day one. You get it by being around long enough, being wrong in front of them once, and coming back with something specific.

If you are reading this deciding whether to hire me for a marketing site, this is the honest pitch: the website is where I start, not where I stop. What I actually want is your business, in detail, and the parts of it that are quietly costing you money.