An AI-first prototype of auto-redlining for construction contracts — built overnight by a designer — that rewrote Document Crunch's roadmap and made agent-first development mandatory engineering practice.
Construction contracts are dense, high-stakes documents. A single missed clause — an indemnification overreach, a missing limitation of liability, a waived right to consequential damages — can expose a contractor to millions in unexpected risk. The industry's solution? Send every contract to outside counsel at $800/hour and wait days for a redlined version to come back.
Document Crunch, a Series C construction technology company, had already built an AI-powered platform for contract review — helping teams identify risks in seconds instead of days. But there was a gap in the workflow customers kept asking about: auto-redlining. Not just flagging risks, but generating the actual suggested contract language and delivering it as a Word document with tracked changes — the format every legal and project team already works in.
The feature was on the roadmap. Someday. It was complex — touching PDF extraction, AI analysis, and Word document generation with OOXML-level track change formatting. A proper engineering effort. Multiple sprints. Cross-functional coordination.
Then a designer decided to build it overnight.
I'm a Staff Product Designer with 15+ years of design experience. I am not a software engineer. I don't write production code for a living. What I have is an AI-native workflow — specifically Claude, running through a messaging app, with the ability to spawn sub-agents that can write, debug, and deploy code autonomously.
The approach was simple in concept, ambitious in scope:
The overnight session started in the evening and ran through the early morning hours. By sunrise, it was live.
The prototype is a full-stack web application with a surprisingly sophisticated pipeline. A user uploads a PDF contract. The system extracts the text using unpdf, a lightweight PDF parsing library. That extracted text is sent to OpenAI's GPT-4o in JSON mode along with a ten-point risk analysis checklist covering common construction contract pitfalls:
The AI returns structured JSON identifying each risky clause, explaining the issue, and proposing revised language. The application then generates a Word document using the docx library — not just inserting comments, but creating real OOXML tracked changes. Deletions are struck through in red. Insertions appear in blue. It looks exactly like a document that's been through a lawyer's review in Microsoft Word.
The Contractor shall defend, indemnify and hold harmless the Owner from and against any and all claims arising out of resulting from the Contractor's negligent acts or omissions in the performance of the Work…
Removed "defend" — shifts burden of counsel to Contractor. Narrowed "arising out of" to "resulting from" to limit scope to actual causation.
From first prompt to deployed application: approximately twelve hours. A single overnight session. One designer. An army of AI agents.
On February 18, 2026, I demoed the prototype to Document Crunch CEO Josh Woolsey and the leadership team.
Josh watched a real contract get uploaded, analyzed, and returned as a Word document with tracked changes. The same workflow that costs customers thousands of dollars in legal fees and days of turnaround — done in seconds by a tool a designer built overnight.
"The fact that Ryan's capable of this either makes Ryan God, or it means that everyone else better get on their shit."— Josh Levy · CEO · Document Crunch
"Ryan, you are my prototype for how this should look moving forward."— Josh Levy · on agent-first development at DC
"This is the compelling event more than others."— Josh Levy · on the strategic significance of auto-redline
Lee, Document Crunch's Head of Engineering, immediately grasped the operational implications. He made three observations that would reshape the company's development philosophy:
"We will die without it."— Lee · Head of Engineering · Document Crunch
The demo didn't just impress leadership. It changed how the company operates.
Document Crunch adopted a new development principle: build agent-first, then GUI. Before committing engineering resources to a feature, prove the concept with AI-assisted prototyping. Lee formalized this as an engineering standard — not a suggestion, a requirement.
Auto-redlining was repositioned from "future consideration" to potentially the biggest feature for Document Crunch 2.0 customers. The overnight prototype proved both technical feasibility and market demand in a single stroke.
The auto-redline prototype didn't just validate one feature. It proved a pattern. If AI can analyze a contract and generate redlined output, the same architecture applies to Notice Builder, COI Review, and Lien Waiver Review — three significant customer value surfaces that went from abstract roadmap items to a concrete, buildable pipeline.
The leadership team saw a designer — not an engineer — build a working, deployed application overnight that addressed one of their customers' biggest pain points. It changed the question from "how many sprints will this take?" to "can someone prototype this by tomorrow?"
With agent-first development now mandatory company practice, I led the redlining discovery as the Principal-level strategic designer for Cycle 5. That work grew into the AI Strike Team, where I architected a three-tier knowledge structure that turns generic language models into domain-fluent AI employees:
The layers compose. The AI doesn't just answer questions — it speaks the company's language, understands its craft, and knows the neighborhood it operates in. The framework shipped to real clients including a custom home builder in Seguin, Texas, paired with a fifteen-task scheduled-automation library.
The overnight redline prototype wasn't the end of a sprint. It was the start of a new way of working.
I'm a designer. I built a full-stack application with AI analysis, PDF processing, and OOXML document generation. The traditional demarcation between "designer" and "engineer" is becoming less about capability and more about preference. AI agents are the great equalizer.
The demo wasn't going to pass a security audit or handle ten thousand concurrent users. It didn't need to. It needed to prove the workflow was possible, the output was useful, and the concept was worth investing in. It did all three in twelve hours.
When a prototype takes months, it's a commitment. When it takes hours, it's an experiment. Document Crunch went from debating whether auto-redlining was feasible to watching it work in real-time — overnight. That speed doesn't just save time; it changes what decisions get made.
Lee's observation — "we will die without it" — isn't hyperbole. Companies that adopt AI-native development workflows will iterate faster, validate sooner, and ship more than those that don't. The gap will only widen.
All the AI tooling in the world doesn't matter if you can't demo something that makes leadership say everyone else better get on their shit. The prototype worked because it solved a real problem ($800/hour legal fees), demonstrated a real workflow (upload → analyze → redline → download), and produced a real artifact (a Word doc with tracked changes). Substance over spectacle.
Twelve hours. One designer. An AI agent orchestrating sub-agents via a messaging app. A deployed prototype that changed a company's roadmap, development philosophy, and competitive strategy.
The construction industry moves slowly by nature — long contracts, longer projects, legacy workflows. Document Crunch exists to accelerate that. On one February night, a designer with AI tools showed the company itself what acceleration really looks like.