A Lawyer, Road Inspector, and Cardiologist Entered a Coding Competition — And Won

On February 20, 2026, Anthropic announced the winners of the “Built with Opus 4.6” hackathon. In this competition where 13,000 people applied and 500 were selected, 277 working software applications were created¹. Over 21 million lines of code were written in 6 days². But when people saw the results announcement, they were more surprised by the winners’ résumés than by the technical sophistication of their creations.

First place went to a personal injury lawyer from California. Second place went to a father who built a programming environment for his 12-year-old daughter. Third place went to a cardiologist from Brussels. Special prizes went to a road inspector from Uganda and a musician from Indonesia. Among the top 5 winning entries, not a single one was made by a traditional software engineer.

The Winning Projects: Software Born from Real-World Experience

1st Place – CrossBeam (Mike Brown)

Mike Brown is a lawyer from California. Not a software engineer, not a computer science major. His friend Cameron was a contractor building ADUs (Accessory Dwelling Units, aka “backyard cottages”), constantly plagued by correction requests during the permit process. Since 2018, California has issued over 429,000 ADU permits, but more than 90% receive correction requests on their first submission. Each correction cycle costs weeks and thousands of dollars, with typical 6-month permit delays generating approximately $30,000 in costs per project².

Brown created CrossBeam. When you upload correction requests and architectural drawings, a California ADU knowledge base consisting of 13 custom “skills” and 28 files kicks into action. Each agent runs in an isolated sandbox, with single analyses taking 10–30 minutes, exceeding serverless function time limits. It was a multi-agent AI system complete with real-time database synchronization through Supabase and parallel skill execution. A lawyer built this in 6 days².

Buena Park Mayor Connor Trout appeared in the demo video, saying: “We need permits for over 3,000 housing units including ADUs by 2029, but we barely hit 100 last year. It’s impossible with current staffing. We need this software.”²

2nd Place – Elisa (Jon McBee)

A visual programming environment for children. Kids drag and drop blocks — describing goals, rules, and skills in natural language — and AI agents generate actual code. While agents work, an educational engine explains what’s happening at the child’s level. He wrote 39,000 lines of code solo in 6 days. The first user was McBee’s 12-year-old daughter¹.

3rd Place – Postvisit.ai (Michal Nedoszytko)

Cardiologist Michal Nedoszytko from Brussels has performed thousands of cardiac procedures. The real difficulty, he says, began “the moment you leave the procedure room.” Patients returned home not understanding their diagnosis, unable to remember what the doctor said². Hearing about the hackathon during his commute, he decided to participate and leveraged Opus 4.6’s million-token context window to create an AI companion that analyzes medical records and files to provide personalized health guidance. He essentially brought AI scribe technology from the doctor’s side to the patient’s table.

Special Prize – Conductr (Asep Bagja Priandana, “Most Creative” Award)

An AI bandmate created by a musician. When you play chords on a MIDI controller, Claude follows in real-time, generating 4-track accompaniment. A C/WASM engine achieved approximately 15ms latency. Rather than AI “generating” music, it was a system where AI “performed” with the musician. Claude Code creator Boris Cherny commented: “The music is great. It’s fire.”²

Special Prize – TARA (Kyeyune Kazibwe, “Keep Thinking” Award)

Built in Uganda, this tool converts dashcam footage into road infrastructure investment reports. Upload footage and get computer vision-based road condition analysis, road segment classification by roughness index, intervention cost estimates, economic analysis including IRR calculations, and equity analysis evaluating who actually benefits. It reduced a process that previously took weeks to 5 hours. It was actually tested on roads under construction in Uganda¹².

”Domain Beats Syntax”

There’s a reason these results weren’t coincidental. One hackathon participant summarized it: “Domain beats syntax. Knowing what to build, and recognizing when the machine gets it wrong — that’s still needed.”²

Lawyer Mike Brown could structure California’s building code system into 13 “skill” documents because reading and cross-referencing legal documents was his profession. The cardiologist could precisely identify patients’ information gaps because he’d experienced “post-procedure room” problems after thousands of procedures. The Ugandan road inspector included IRR calculations and equity analysis in the requirements because he was someone who had done this process manually.

AI replaced the ability to write code. But it didn’t replace the ability to define problems. Rather, the value of that ability rose dramatically. Boris Cherny’s evaluation of first-place CrossBeam captures this: “I liked the user focus. Go to users and ask what they want, then build it. Go to other users, ask again, build again. That’s how the Claude Code team builds software too.”²

Vibe Coding Becomes Reality

In February 2025, OpenAI co-founder and former Tesla AI head Andrej Karpathy wrote on X: “There is a new kind of coding. I call it ‘vibe coding’. You just go full send on vibes, embrace exponential progress, and forget that code even exists.”³ This single sentence spawned a concept, got its own Wikipedia entry⁴, and prompted IBM to create an official explainer page⁵.

In December 2025, reflecting on the year, Karpathy wrote that vibe coding had produced a “free and disposable” new type of code. He added a prediction: vibe coding would “terraform software and alter job descriptions.”⁶ The Built with Opus 4.6 hackathon results show this prediction is already becoming reality.

What’s important is that these hackathon winners didn’t create “toys.” CrossBeam was software that mayors said they needed. Postvisit.ai was at a level that could be integrated into actual clinical workflows. TARA was tested on real construction sites. Conductr was a real-time music system with 15ms latency. These were closer to early products than prototypes.

The Great Transformation of Software Development

Simply summarizing this phenomenon as “non-developers can now code” is only half right. A more accurate description: the most expensive bottleneck in software development has shifted from ‘implementation’ to ‘problem definition’.

The U.S. Bureau of Labor Statistics (BLS) projected 15% growth in software developer employment between 2024–2034⁷. However, this figure was calculated before the rapid advancement of AI coding tools, and real-world signals point in different directions. According to TrueUp data, there were 783 tech company layoffs in 2025 affecting 245,953 people. In 2026, by late February alone, 112 layoffs have affected 40,132 people⁸. While not all layoffs are due to AI, there’s growing recognition that AI increases coding productivity, allowing the same output with fewer people.

Simultaneously, new types of software are emerging explosively. Boris Cherny called FenceFlow — an automated fencing tournament tool that won a special prize — “what the future looks like.” He meant “custom software for communities that are too niche for commercial products to ever exist.”² AI analysis tools for the fencing world. Specialized systems for California ADU permit corrections. Economic analysis pipelines for Ugandan road infrastructure. Such software couldn’t have been created in the existing software market. The market size didn’t justify development costs.

AI coding tools changed the ‘development cost’ term in this equation. Six days, one person, API credits. If this alone can create software that solves real problems, the range of “problems worth making software for” expands dramatically. Consequently, the total amount of software is likely to explode rather than shrink. However, the profile of people making that software is changing.

Who’s Threatened, Who’s Rising

Pushing this hackathon’s lessons to extremes yields two scenarios.

In the pessimistic scenario, the value of developers who “just implement” plummets. The ability to receive requirements and convert them to code — traditionally a core competency of junior developers — becomes commoditized by AI, and as barriers to entry lower, the scarcity of existing practitioners decreases too.

In the optimistic scenario, the boundary between domain experts and developers blurs, expanding the very definition of “technical problem-solving ability.” In a world where lawyers build legal systems, doctors build medical systems, and inspectors build infrastructure systems, “developer” becomes not a specific job title but a capability embedded in all professions.

Reality will probably be somewhere in between. But what the Built with Opus 4.6 hackathon clearly showed is that the optimistic scenario is no longer theory but something already happening. In a competition where 13,000 applied and 500 were selected, the people who created the most impressive software weren’t people whose job was making software. Their weapon wasn’t coding skills but field experience.

Karpathy said vibe coding would “alter job descriptions.” This hackathon was the first snapshot of that change. A lawyer, road inspector, and cardiologist entered a coding competition. And they won. This isn’t the setup for a joke. It’s the opening act of a new reality.