Tri-state area HVAC & generator company
Generator proposals in minutes, not hours
Enter the job, get a priced, branded proposal the customer can sign online.
The problem
Where the time was going.
Every quote was rebuilt by hand: a spreadsheet for the math, a Word template for the document, and the owner's memory for everything in between. Each one took hours.
Three years of past proposals sat in about 45 customer folders that nobody could search. Installs were quoted as one lump sum, so there was nothing to compare a new job against, and the generator size was often picked by feel.
While a quote was being put together, bids were missed and some jobs went to competitors who answered faster.
What we built
The fix.
A web app where the team describes a job whichever way is easiest: the site survey form, a short chat or dictation, job photos, an uploaded PDF, or a public bid package of 100+ pages. Techs can start an estimate from their phone on site.
We read and structured 121 of their past proposals. Installs are now split into materials, mechanical, and electrical, and each part, like a crane lift, a new pad, or a gas run, is priced from the same element on similar past jobs. Those jobs are cited on every line.
Built-in sizing uses NEC load calculations to pick the smallest unit that covers the demand. A second-reader check flags easy-to-miss problems before export, like a transfer switch smaller than the service or a diesel unit with no fuel tank.
Every new proposal goes back in. Won jobs rank first as comparables, and when the team keeps making the same price edit, the tool proposes a new pricing rule for the owner to approve.
The finished proposal exports to Word or PDF in their own layout and goes to the customer with a signing link. They draw or type a signature, and a signed copy comes back automatically.
Results
What changed.
Hours by hand became minutes. In an accuracy check, a 20 kW residential job priced by the tool landed on the exact total of a real proposal the company had won, and the equipment matched the owner's own workbook to the cent.
How it works
Stack and approach, in plain language.
- Next.js
- Neon Postgres
- Claude (Anthropic API)
- docx, pdf-lib, and pdf.js
- Resend
- Vercel
- The job comes in through a form, a chat, photos, or a document. Claude reads surveys and bid packages, runs the chat, reviews the worksheet like a second estimator, and writes the scope of work. Large PDFs are read right in the browser.
- The AI never sets a price. Every line is tagged with where its number came from: the price list, a past job, a public bid, an assumption, or an allowance. Export stays blocked until a person signs off on anything not backed by hard data.
- When the manufacturer's price sheet changes each year, an admin uploads the new one, reviews a line-by-line diff, and applies it. No re-keying.
- Proposals are generated as Word or PDF, emailed with a private signing link, and signed online. Won and lost outcomes feed back in, so the next quote has more to go on.
Next step
Got a similar problem?
Thirty minutes, no cost, no pitch deck. Tell us what is eating your time and we will tell you what we would build.
More work
See everythingHVAC & generator company, NJ
An AI receptionist that picks up every callHVAC & generator company, NJ
A QA/QC checklist that checks itself against Housecall ProTri-state area HVAC & generator company
An AI operations manager that watches 18 systems at once