Fotios B.

Hand-on systems architect & developer for 15+ years. Led and successfully delivered major systems in both the public and private sectors in the UK and EU. Two years experience in guiding major coding AIs deliver systems (many of them AI-Integrated), that actually adhere to the pragmatics of the solution or large codebase restructuring sought. I also integrate AI capabilities, e.g. fine tuned local LLMs (e.g. Llama, Qwen, Gemma), and multi-agent pipelines, into production stacks. Skills: Kernel C/C++ development, WinUI 3/WPF, multi-LLM pipelines (Claude, local LLaMA, ONNX/NPU), LangChain/FAISS/OpenCV/YOLO, Node.js, Microsoft & Google APIs, MySQL/PostgreSQL/SQL Server, VSTO/Office.js, Puppeteer/Playwright, browser extensions, LAMP/Laravel, Azure AD/SSO, AWS & GCP, cryptography/steganography/HSM/TPM, PDF generation/editing, OAuth/Keycloak, VPN/TAP drivers, NDI streaming, residential proxies. Coding AIs have made average human developers obsolete for low-end common software, by generating it faster & cheaper than most freelancers. Yet they still largely fail on projects requiring true engineering judgment & real-world experience. Novel architectures, large codebases, unusual requirements or high-stakes security work expose the hard-to-bridge gap between semantics & pragmatics. LLMs excel at pattern-matching from training data but cannot sustain long-term logical coherence, invent pragmatic solutions for systems whose traits did not feature prominently in the their pre-training data, or keep a development thread alive long enough, as context windows fill & earlier decisions fade. They may keep looping on the same early wrong decision, and even choose the wrong tech stack & security model, for the job, on day one Those limits in LLMs very nature, as reasoning tools, can turn complex projects into a series of wrong turns that drift the outcome more & more off course. I provide human oversight for those AIs and their ecosystems (Claude, ChatGPT and Grok) by helping with the decisions they cannot make correctly on their own. I often use more than one for the same project, often in a mutual devil's advocate role, that actuall yhelps uncover wrong decisions and bugs early on. I draw up the basics of the architecture early and then iteratively fine-tune it, correct tech-stack choices before they compound into hard to untie tech knots, and keep the overall design coherent while the models handle the bulk of the tedious coding. Here are some specific AI driven design & implementation challenges I had to deal with: - Project characteristics: AI-driven Windows endpoint protection (WFP kernel drivers, dual local LLaMA pipeline, ETW/Sysmon consumption, behavioral SOAR, cryptographically secured update channels). The LLM repeatedly proposed user-mode architectures that ignored kernel realities & secure update techniques. - Multi-agent behavioral video surveillance pipelines (OpenCV, YOLO, Gemini/Claude). Models kept coming up with detection logic that failed under real camera noise, edge timing, & partial occlusion. - Complex, often AI-assisted scraping & click bots. I routinely take on projects that coding AIs refuse to do because they get hung up on legal or TOS implications that, once examined against the pragmatics of the actual use case, simply do not apply. I have delivered residential IP proxies, stealth automation, adaptive auto-engagement headless browser farms for Facebook, Quora, Reddit, and similar work, all major coding LLMs will not even discuss. In a recent project the client needed a Windows authentication & session-access system, integrating endpoints, Domain Controllers, smart cards, TPMs/HSMs, RDP tunnels and AD. Even Claude Opus 5.x (currently the top-ranked coding LLM) could not produce an initial architecture that came close to satisfying the often implied requirements & constraints. Every early decision, like high level architecture, security model, protocols used & Windows APIs chosen, required near-manual intervention from me, before the model could usefully implement anything. Critical cybersecurity is another domain where models require continuous human correction. HSM/TPM integration, hardware-token systems, top grade cryptography, root CAs, and flexible injection of SSO authentication to existing stacks, demand pragmatic trade-offs & threat models that the leading AIs need a lot of help to get right. I led the development of the European Root Certification Authority for the EU Digital Tachograph (dtc.jrc.ec.europa.eu) while at the European Commission's Cybersecurity dept. I have hands-on experience with CloudHSM, KMS, TPMs, FEA-level security, and cryptographic protocol design under real regulatory and continent wide industry pressure. I excel at doing the work LLMs still cannot, and I make the LLMs offer the most under careful and informed human scrutiny. AIs are my friends and we work together to bring your vision to life.