Rebuilding a Founder's Vision Into a Real Product

A stylist who works fashion week and major brands had an idea for how his industry should run, and he'd already proven it himself, teaching himself just enough to build a working version alone. Rootstrap came in to take that vision the rest of the way: from something that worked in a demo to something that could carry his business.

“On an active job I’d spend 30 to 60 minutes a day logging and organizing receipts, more when returns stacked up. With Archivo it’s a few seconds per receipt. A big job’s wrap report used to take me one to two full days of matching receipts, returns and refunds. Now it takes minutes, because the report builds itself as the job goes.”

David Gangel · Founder, Archivo

The Problem

A stylist advances tens of thousands of dollars of their own money on every production. A lost receipt used to mean $3,000 gone for good, out of the stylist's pocket, not the production's. That doesn't happen anymore. Industrywide, $850M to $1.5B in samples leave each year and never come back, and that's the problem Archivo exists to close. The founder set one bar for the product from day one: the numbers have to be 99.99% accurate.

About Archivo

Archivo helps stylists manage the parts of the job nobody sees: tracking receipts, chasing down returns on borrowed garments, putting together wrap reports for agencies and clients when a project wraps, all backed by an in-app AI assistant. The founder built the first version himself with AI no-code tools (Lovable, Cursor, Supabase, Capgo), enough to prove the idea was real, but not enough to run on.

Powered by Claude

Two distinct uses of Claude sit at the core of the product. Haiku handles multimodal structured data extraction, parsing photos and PDFs of receipts and rental agreements into structured records. Sonnet powers Arco, the in-app conversational assistant, with live tool access to the application database so users can query and act on their own data in natural language.

One AI provider does nine distinct jobs: receipt extraction, purchase/return classification, bank statement parsing, rental agreement extraction, garment ID, labeling, drafting, the in-app assistant, and the daily brief. It replaced two separate OCR vendors. It never guesses with someone else's money: auto-reconciles only above a 70/100 match, asks for a quick human confirmation between 40 and 70, and touches nothing below 40. A failed extraction never gets saved as a silent zero. Bank statements are never stored.

Starting With Intent, Not Assumptions

Before touching the code, Rootstrap sat down with what the founder actually wanted the product to become, and audited what had actually been built against it. The gap was real: working features sat on a foundation with no safety net.

• Type-checking was disabled across the entire codebase
• Automated test coverage was effectively zero
• Deployment ran through a single script, executed by hand, with no staging environment behind it

Getting the Ground Truth Right

That audit became the reference point for everything after. Rather than guess at what "done" meant, Rootstrap built a living, human-reviewed record of what the client had actually asked for and approved, keeping current as requirements shifted.

• Used it to sequence the rebuild itself: foundation first, features second
• Required every AI agent in the delivery pipeline, QA included, to check against it before acting
• Used it to catch a real case where a legitimate client-requested change had been misflagged as a regression

Building It Back Up

With that ground truth in place, the team rebuilt the engineering foundation, then layered in what the prototype never had.

• Installed a CI/CD pipeline and a staging environment
• Brought strict typing and test coverage to the app's most critical paths
• Closed the security gaps the audit had surfaced
• Delivered payments integration, shipping, database-enforced role permissions, and GDPR/CCPA compliance readiness

Verifying It Held

The AI QA layer checked its own work against that same ground truth before flagging anything as broken, catching the difference between an actual regression and a decision the client had already made. It's a small thing that mattered: it kept the team from chasing bugs that weren't bugs, and kept the QA process itself trustworthy at production speed rather than just fast.

What It Proved Out

A stable, working baseline came together in the first three weeks of a ten-week first phase, built on a foundation that could actually hold new features without breaking. The client trusted the rebuild enough to bring Rootstrap a second phase, and a new line of business, before that first phase had even wrapped. That same engine now powers a second vertical, 23 manufacturing capabilities from fiber to factory exit, with no rewrite required.