Within the Writer Highlight sequence, TDS Editors chat with members of our neighborhood about their profession path in knowledge science and AI, their writing, and their sources of inspiration. In the present day, we’re thrilled to share our dialog with Mike Huls.
Mike is a tech lead who works on the intersection of knowledge engineering, AI, and structure, serving to organizations flip advanced knowledge landscapes into dependable, usable techniques. With a robust full-stack background, he designs end-to-end options that stability technical depth with enterprise worth. Alongside shopper work, he builds and shares sensible instruments and insights on knowledge platforms, AI techniques, and scalable architectures.
Do you see your self as a full-stack developer? How does your expertise throughout the entire stack (from frontend to database) change the way you view the information scientist position?
I do, however not within the sense of personally constructing each layer. For me, full-stack means understanding how architectural selections at one layer form system habits, threat and price over time. That perspective is crucial when designing techniques that must survive change.
This angle additionally influences how I view the information scientist position. Fashions created in notebooks are solely the start. Actual worth emerges when these fashions are embedded in manufacturing techniques with correct knowledge pipelines, APIs, governance, and user-facing interfaces. Knowledge science turns into impactful when it’s handled as a core half of a bigger system, not as an remoted exercise.
You cowl a variety of matters. How do you determine what to concentrate on subsequent, and the way are you aware when a brand new matter is value exploring?
I are likely to comply with recurring friction. After I see a number of groups wrestle with the identical issues, whether or not technical or organizational, I take that as a sign that the difficulty is structural somewhat than particular person, and price addressing on the architectural or course of stage.
I additionally intentionally experiment with new applied sciences, not for novelty, however to know their trade-offs. A subject turns into value writing about when it both solves an actual drawback I’m at present dealing with or reveals dangers that aren’t but broadly understood. Lastly, I write about matters I personally discover attention-grabbing and price exploring, as a result of sustained curiosity is what permits me to go deep.
You’ve written about LangGraph, MCP, and self-hosted brokers. What’s the largest false impression you assume individuals have about AI brokers at the moment?
Brokers are genuinely highly effective and open up new potentialities. The misunderstanding is that they’re easy. It’s straightforward at the moment to assemble cloud infrastructure, join an agent framework, and produce one thing that seems to work. That accessibility is effective, but it surely masks quite a lot of complexity.
As soon as brokers transfer past demos, the actual challenges floor. State administration, permissions, value management, observability, and failure dealing with are sometimes underestimated. With out clear boundaries and possession, brokers turn out to be unpredictable, costly, and dangerous to function. They don’t seem to be simply prompts with instruments; they’re long-lived software program techniques and have to be engineered and operated accordingly.
In your article on Layered Structure, you point out that including options can usually really feel like “open-heart surgical procedure.” For a newbie or a small knowledge crew trying to keep away from this, what’s your key recommendation on establishing an structure?
“The one fixed is change” is a cliché for a very good motive so optimize for change somewhat than for preliminary supply velocity. Even a minimal type of layered considering helps: separating area logic, software stream, and infrastructure issues.
The aim shouldn’t be architectural perfection on day one or excellent categorization. It’s about creating clear boundaries that enable the system to evolve with out fixed rewrites. Small upfront self-discipline pays off considerably as techniques develop.
You’ve benchmarked PostgreSQL insert methods and famous that “sooner shouldn’t be all the time higher.” In a manufacturing ML pipeline, what’s a state of affairs the place you’ll intentionally select a slower, safer insertion technique?
When correctness, traceability, and recoverability matter greater than uncooked throughput. In lots of pipelines, decreasing runtime by a number of seconds affords little profit in comparison with the chance launched by weaker ensures.
For instance, pipelines that feed regulatory reporting, monetary decision-making, or long-lived coaching datasets profit from transactional security and express validation. Silent knowledge corruption is way extra pricey than accepting modest efficiency trade-offs, particularly when knowledge turns into a long-term asset others will construct on..
In your Private, Agentic Assistants article, you constructed a 100% personal, self-hosted platform. Why was avoiding “token prices” and “privateness leaks” extra necessary to you than utilizing a extra highly effective, cloud-based LLM?
In my day by day work I’ve skilled that trusting a system is prime to system adoption. Token prices, opaque knowledge flows, and exterior dependencies subtly affect how techniques are used and perceived.
I additionally made a aware selection to not route my private or delicate knowledge by exterior cloud suppliers since there are restricted ensures on how knowledge is dealt with over time. By holding the system self-hosted, I might design an assistant that’s predictable, auditable, and aligned with European privateness expectations. Customers have full management over what the assistant has entry to and this lowers the barrier for utilizing the assistant.
Lastly, not each use case requires the most important or most costly mannequin. By decoupling the system from a single supplier, customers can select the mannequin that most closely fits their necessities, balancing functionality, value, and threat.
How do you see the day-to-day work of a knowledge skilled altering in 2026?
Regardless of widespread stereotypes, knowledge and software program engineering are extremely social professions. I strongly consider that probably the most vital a part of the work occurs earlier than writing code: aligning with stakeholders, understanding the issue house, and designing options that match current techniques and groups.
This upfront work turns into much more necessary as agent-assisted improvement accelerates implementation. With out clear objectives, context, and constraints, brokers amplify confusion somewhat than productiveness.
In 2026, knowledge professionals will spend extra time shaping techniques, defining boundaries, validating assumptions, and guaranteeing accountable habits in manufacturing environments.
Trying forward at the remainder of 2026, what huge matters will outline the 12 months for knowledge professionals, in your opinion? Why?
Generative AI and agent-based techniques will proceed to develop, however the larger shift is their maturation into first-class manufacturing techniques somewhat than experiments.
That transition will depend on reliable, high-quality, accessible knowledge and strong engineering practices. Consequently, full-stack considering and system-level design will turn out to be more and more necessary for organizations that need to apply AI responsibly and at scale.
To study extra about Mike’s work and keep up-to-date together with his newest articles, you possibly can comply with him on TDS or LinkedIn.
