AI Adoption: Our Stance on Local LLMs and Workflows

Why We Replaced Paid Cloud APIs with On-Premise AI for Daily Development Workflows

In-house, our AI journey has been a continuous exercise in cognitive dissonance. On one hand, the jump from mediocre LLMs just a couple of years ago compared to today’s local models has been incredible. For teams like ours paying close attention, the sheer viability of local models has turned what used to be a tinkerer’s hobby into serious production-grade tooling.

Recently, we collaborated with Goodfirms to discuss how AI is actively steering our internal workflows and what adaptations we have had to make. We joined them for an in-depth video conversation to share our honest stance, not from the vantage point of AI hype, but from the daily trenches of software workarounds and delivery. You can check out the full discussion in our collaborative video here.

By starting the video, you agree to YouTube's privacy policy. Open YouTube

A sincere thank you to the Goodfirms team for inviting us into such a timely, grounded conversation about the real complexities of modern engineering.

We shared our hands-on in-house experience using LLMs for offline use, which has turned out to be worth sparing every penny. And when asked about day-to-day productivity, the verdict remains clear that my local LLM replaced ChatGPT for most of my daily work.

Running local LLMs on my own machine for unlimited coding and writing tasks has completely moved my expectation of what's possible, especially as token costs spiked. On the other hand, it has exposed a huge, flawed misconception across the industry: the idea that anybody can build production-grade software without developers. The unexpected challenge is that AI makes generating code deceptively easy.

— Chris Wolf, CEO & Senior PHP Developer at Laramate GmbH

This naturally prompts the question, will AI replace programmers? The short answer is no. What we are seeing instead is a fool's gold rush where anybody can prompt a prototype, but very few can actually deploy a secure, maintainable, scalable solution.

A pile of unverified AI code is not a sustainable application; generating syntax represents only a fraction of the actual development effort. While writing boilerplate has gotten exponentially faster, solid system architecture, data compliance, and human design thinking remain non-negotiable. AI accelerates iteration and execution, but it cannot substitute for deep foundational judgment.

That is why keeping the human in the loop is essential. It also explains why you need to run a local model instead of routing proprietary business logic through third-party cloud APIs. If you ask us why we use local models the answer usually comes down to privacy, cost control, and protection of intellectual property.

Our own experiences with local models for coding have proven that running local models is good now, and not just "good enough" but reliably performant. With modern quantisation techniques and lean parameter architectures, local models are actually good now at handling heavy contextual workloads broken down into discrete tasks and automating repetitive tasks on consumer hardware.

Local models force discipline. Since the local model has to deal with the hardware limitations of a consumer‑grade laptop, it cannot one‑shot a solution of that size. Tasks that are supposed to be handed over to a model, need to be small, clearly defined and contain relevant context and skills. This costs a significant chunk of the available context. 

It leads to more human involvement in the overall process because under these circumstances, one can keep up.

— Chris Wolf, CEO & Senior PHP Developer at Laramate GmbH

The trick lies in figuring out which local models are best for which tasks; using compact, specialised coder models for tight autocomplete loops, and keeping larger instruct models primed for architectural reviews or technical writing.

So, are local models the future of AI? For teams that value data sovereignty, predictability, and deep control over their development environments, the answer is an emphatic yes. The ability to use AI models locally gives developers an unrestricted, zero-marginal-cost copilot that lives strictly on their terms. But as we discussed with Goodfirms, tooling only magnifies existing skills.

What businesses need today is experienced guidance on how to avoid AI-generated pitfalls because you can only eliminate technical debt when you actually know how to identify it.