Reading napkin orders into the ordering system, with a human in the loop

Ventura, Summer 2025, later in Y Combinator (Winter 2026)

Guessing wrong costs a customer. I joined Ventura as its second person. By summer’s end, 15 industrial distributors were paying for our agent.

about $600,000 closed in three months, starting from an idea

Signed contracts only.

Role
Co-⁠founder, product and go-⁠to-⁠market
Team
Two founders, then a chief technology officer I brought on and a small engineering team
Timeline
June to August 2025
What changed
From retyping orders all day to checking what the agent flags

In short

  • A rep types every order, from faxed purchase orders to the odd napkin, line by line into the distributor’s ordering system, its enterprise resource planning (ERP) software.
  • Our agent matches each line to the catalog and flags the ones it isn’t sure of.
  • The sales calls decided what we built: the one job distributors described.
The Clean-Data demo: a forwarded text message read into 20 order lines, with 19 matched to the catalog and 1 flagged for a person to check before anything goes to the ordering system

The numbers

more than $1 millionqualified pipelineOpportunities past discovery, not total outreach.
10%of cold outreach answeredCompared with the usual 1 to 2 percent.

My part

  1. Decidedthat nothing would reach the ERP until a person had resolved every line the agent flagged.
  2. Ranevery sales call, the pricing and the business plan as the de facto product manager.
  3. Builtthe agent, from reading the order to syncing it to the ERP.
  4. Builtthe public demo of the review workflow, so anyone could see it without a sales call.Clean-Data

Check the two lines the machine isn’t sure of.

In

  1. Emailneed 20 of the 3/4 ball valves brass + 10x 1/2 npt tees
  2. Faxed purchase orderPO 58812: 50 HEX BOLT 3/8-16 x 2 ZN
  3. Counter slip6 x 2" pvc cplg, 2 x pipe dope

Matched to the catalog

  1. Ball valve, 3/4", brass, full port× 2097% sure
  2. Tee, 1/2" NPT, black iron× 1088% sure
  3. Hex bolt, 3/8"-16 × 2", zinc× 5099% sure

For a person to check

Two lines need a person: pick one to check it.

The orders and parts are made up. The order reader is a perforating machine from the Smithsonian’s National Postal Museum.

Clips

In this edge case, a forwarded text message turns into 20 order lines, with 9 typos fixed, 28 abbreviations decoded and one line flagged for a person. Clean-Data still has Ventura’s name on it. At the end the flagged line is skipped. The demo isn’t connected to a real ERP, so Sent to ERP only marks the order approved.

The problem

An inside-sales rep spends most of the day retyping orders, and every line is a chance to get a part number or quantity wrong. The orders are a mess: an email that says “blk iron niples 3/4”, or a purchase order scanned at an angle. The rep knows the catalog well enough to decode all of it. That’s why the job is so hard to hand to software.

From an idea to 15 paying distributors

We had one summer. I owned the outreach, the calls and the prototypes. Plenty of ideas sounded good from the engineering side, including my co-founder’s, but distributors could tell in one conversation that they wouldn’t help. I set those ideas aside and kept the job distributors described: getting the order into the ERP the way their best rep would.

By the end, the distributors thanked us. So did the reps, who had spent their days entering orders and got that time back.

Decision log

Chosen

to test every sentence of our cold outreach, the subject line and the channel.

Rejected

sending one good pitch to everyone.

Why

we didn’t know the buyer yet.

Result

10 percent of the people we reached answered, compared with the usual 1 to 2 percent.

Chosen

to lock “Send to ERP” until a person has resolved every flagged line.

Rejected

sending orders automatically.

Why

a wrong part that ships means a returned pallet and, often enough, a lost account. A flagged line costs a rep about ten seconds.

Result

the review step became the demo.

The bottom of an order review in Clean-Data. The lithium grease, cut-off wheel and pipe nipple lines each have a catalog part number and a match score from 94 to 98 percent. The line camlock fittings 2in alum has a red Unknown tag at $0.00, with skip and include buttons. Below the table, the order total is $10,312.85 and the send button reads Resolve 1 flagged item
From Clean-Data, my public rebuild of the review workflow, in October 2026. The demo couldn’t match 40 camlock fittings to the catalog, so it marked that line Unknown. Until someone skips or includes it, the send button reads Resolve 1 flagged item.
Chosen

to show where each line came from: hover over it to see the original words.

Rejected

a clean table.

Why

reps needed to check the agent’s work.

Result

in the first month, reps came to trust the agent by checking those sources and finding them right.

Rows from Clean-Data’s review of a sample email from a contractor. A tooltip labeled Original Input sits over the line Ball Valve 3/4 inch NPT, Brass, Full Port, 200 valves at a 93 percent match, and quotes the sentence it came from: Also need 200 brass ball valves 3/4 full port and 175 bronze gate valves 1 inch.
From Clean-Data in October 2026. Hovering over the ball valve line shows the sentence it came from, in a sample email from a contractor. The same sentence asked for 175 bronze gate valves, which became a line of their own.
Chosen

to sell the review workflow.

Rejected

leading with the model’s accuracy numbers.

Why

every buyer we met had heard accuracy claims. None had seen a review step designed around their own people.

Result

about $600,000 closed in three months.

Chosen

to leave pricing rules and live inventory checks out.

Rejected

building a complete order-to-cash flow.

Why

the whole promise was matching lines correctly, with a person signing off on the uncertain ones.

Result

the first version did one thing well enough that customers paid for it.

What I would do differently

I’d go in person and spend a day shadowing a distributor’s order desk, instead of relying on how people described their process. I’d also plan the summer better. The testing paid off, but it pushed the building into the last few weeks, and I pulled one all-nighter after another. I should have asked for help sooner.

What I would do next

I’d measure the two numbers a distributor cares about, before and after: the minutes from an order arriving to its entry in the ERP, and wrong part numbers per hundred lines. We had testimonials but not those numbers.

A postcard from San Francisco.

Postmarked 1951. Photograph by Hope Moore (public domain). Click or tap the card to read the note on the back.

Open the live app

Try it: run the demo, resolve what it flags, and send the order to a mock ERP.

Go deeper

A Golden Gate Bridge tower rising out of fog, its red cables running down into the gray
The Golden Gate Bridge in fog. Photo by Carol M. Highsmith, 2012, Library of Congress.