Matt Shumer wrote recently that "something big is happening" in software development. He described the inflection point where AI models are no longer just assistants but genuine collaborators that can build production systems alongside you and the nature of building is fundamentally changing.
I agree with his observations and the timing came even sooner for me than it did for him. Our inflection point at my ventures came months earlier, and what we've built since then is proof that this shift is real, practical, and available to any small team willing to work with the tools.
The Lexacon.AI story: three phases of building
In April 2025 I co-founded Lexacon.AI, an AI document intelligence platform for construction project administrators and commercial manager. The company started through the Antler Entrepreneur-in-Residence program. The first month was pure customer validation: 70+ interviews, rapid prototyping with Lovable, testing hypotheses against real client needs.
Then came the build.
Phase 1: Customer validation (Apr-Jun 2025) 70+ customer interviews. Rapid prototyping with Lovable. Product-market signal established. Our first encounter with using AI to create applications. Lovable required immense number of iterations on simple features but was faster than hiring someone and we had no cash anyways.
Phase 2: Traditional development (Jun-Oct 2025) Hand coding with Cursor, ChatGPT for code examples, adapting by hand. Six months of work produced roughly 1/3 of the current system with a lot of learning the hard way since none of us had built a modern AI application before.
Phase 3: The Claude Code switch (Oct 2025 - present) Completely switched to Claude Code. Rebuilt the entire 6 months of work in approximately 1 month with a new, scalable architecture. Then built 2x more on top. The full 12-month roadmap, completed in 3 months.
Read that again: six months of traditional development produced one-third of the system. Three months with Claude Code completed the full twelve-month roadmap and then some.
The Opus 4.5 inflection point
The shift wasn't gradual. There was a specific moment.
Before Opus 4.5, AI-assisted coding was promising but inconsistent. Sonnet 4.0 was capable but made too many mistakes to rely on. Opus 4.1 delivered good quality but burned through credits so fast that features took weeks to ship to the quality bar we needed.
Opus 4.5 changed the equation. Quality, speed, and cost all improved simultaneously. Features that took weeks now shipped in days, roughly an 80% reduction in build time. And the output quality was higher, not lower.
That's when it became clear: this isn't an experiment anymore. Any small team with the right approach can now build production AI systems.
It isn't vibe coding anymore
- 20,000+ documents processed
- 80,000+ embedded chunks
- MCP self-service for clients
Customers access Lexacon's tools directly through MCP, plugging in Claude Code or other AI assistants for self-service document processing, search, and analysis. The platform that took years on the original trajectory is live, serving real customers, with a team of two. Our stack passes security assessments and can be deployed in our customer's virtual private clouds. We have a robust CI/CD pipeline and full automated test coverage. All major functions in the tool are accessible via MCP: ingest thousands of documents, classify them, extract the text, extract key data, use our advance AI based retrieval and synthesis. That means they can use their data and make their own analysis without us having to build them new features. Removing us as a barrier to faster value.
Building the consulting business on the same stack
After building Lexacon with these tools, the next question was obvious: can we help other teams do the same thing?
Black Hills Labs (BHL) runs its entire consulting operation on the same stack: Claude Code for AI-assisted operations, n8n for workflow automation, NocoDB for database and pipeline management. Total infrastructure cost: approximately $5/month.
120+ companies researched. 500+ contacts managed. Personalized outreach drafted in batches. Lead scoring, proposal generation, session preparation, all systematized with CLAUDE.md rules that encode business logic. Sales dashboarding, Google Ads optimization, a control centre for all of our client efforts. The tools I sell to clients are the tools I run my business on.
What this means for SMEs and startups
There's a gap in the market. On one side: free tutorials and YouTube courses. On the other: enterprise consultants charging $100K+ per engagement. In between, where most SMEs and startups actually live, there's almost nothing.
That gap is closing. The same tools that let a two-person team build a production AI platform are available to any business willing to invest a focused sprint. Not $100K and six months. A few thousand dollars and a few weeks. The technology has crossed the threshold. The question is who moves first.
Introducing: Building with Claude
This article launches a series where I document exactly how we build with Claude Code, both the product (Lexacon.AI) and the business (Black Hills Labs). A third stealth venture is coming and I'll walk you through that too once we're out.
Two tracks:
- The Lexacon track: how we built a production AI platform. Technical, metrics-heavy, architecture decisions.
- The BHL track: how we run a consulting business. Operational, process-focused, with the actual rules and workflows we use.
Not theory. Not predictions. Real production systems with real metrics, documented as we go.
If there's a topic you'd like us to cover in this series, reach out and let us know.
Want to know where you stand?
Take our AI readiness quiz. 10 questions, 2 minutes. You'll get a personalized recommendation for where to start with automation.