# Dragos Bilaniuc

> Independent AI Reliability Reviews · Software Architect. I find where AI products break, and the design decisions that let it break. Independent reviews of LLM products: verified findings, architecture review, system-level fix plan.

- Based in Cluj-Napoca, Romania; works remotely across EU and US time zones.
- Offer: The AI Reliability Review — an independent, fixed-scope review of LLM products. Deliverables: Failure inventory, Architecture findings, Ranked fix roadmap, Starter regression set, Executive readout.
- Engagements run through Luckylabs Software SRL.
- Contact: dragos@dbln.me

## Pages

- [Home](https://dbln.me/): positioning, the review offer, the method, selected work, testimonials, writing
- [Work](https://dbln.me/work): 6 case studies in depth (situation, the expensive-to-reverse decision, approach, outcomes) — Pie Insurance, Bullseye Web3 Studio, Parentool, Glede, Reach Finance, Equinet
- [Writing](https://dbln.me/blog): all articles; RSS at https://dbln.me/feed.xml
- [Privacy Notice](https://dbln.me/privacy): what the site collects (no cookies, cookieless analytics, Cal.com only on request)

## Writing

Every article is also available as raw markdown: append ".md" to its URL.

- [Putting the trace before the loop](https://dbln.me/blog/building-an-ai-agent-observability-first): Observability is the bedrock of reliability, so I designed and implemented the full tracing layer for an AI agent before writing a single line of the agent loop. Three benefits, two drawbacks, and the span tree that turned out to be the loop's control flow.
- [Did it check, or did it guess? Testing a tool whose output is written by an LLM](https://dbln.me/blog/testing-ai-tools): My first AI tool worked, or at least the output looked right when I ran it. Then I pictured a client asking what makes it trustworthy: did it check, or did it guess? This is how I built a real answer, and how designing the output for testing made the tool harder to lie with.
- [I designed an AI auth-auditing tool around honesty. On its first real run, it confidently lied to me.](https://dbln.me/blog/designing-honest-ai-tools): The tool was built to detect what it can verify and refuse to guess at the rest. On its first real run it confidently got the basics wrong. The fix had little to do with the prompt and almost everything to do with the structure around it.
- [Keep the steering wheel: 3 ways to future-proof your auth](https://dbln.me/blog/future-proof-auth-architecture): Vendor auth libraries quietly become load-bearing architecture nobody chose. Three practices that keep an auth system migratable — and why coding agents make the resilient version nearly free to build.
- [Securing auth in a large-scale production system: three industry-standard architectures — and why none survived a closer look](https://dbln.me/blog/securing-auth-large-scale-production-system): Evaluating the auth layer for a Partner Portal of 100+ microservices. Three textbook architectures each looked right on the whiteboard and fell apart on contact with the constraints — here's the one that survived.
- [Sketching out the extremes: an approach to designing software architectures in highly unpredictable projects](https://dbln.me/blog/software-architecture-in-unpredictable-projects): When requirements are genuinely unknowable, picking an architecture up front is a guess. Sketching the two extreme designs first turns that guess into a decision you can defend.

## Profiles

- [GitHub](https://github.com/dragosbln)
- [LinkedIn](https://www.linkedin.com/in/dragosbln)
- [dev.to](https://dev.to/dragosbln)