bence@portfolio:~$ ./intro --short
I build internal tools and process automation, working with AI agents.
Full-stack background (JavaScript, C#) · Power Platform in production · ML validation. Budapest, in Hungarian and English.
Replaced equipment requests handled via Teams messages and manual Excel: requests, issuing, returns, receiving and inventory in one Power Apps system, traceable down to serial number. Live in production; the stock overview is still in development.
Power Apps (Canvas app) · Power Automate · SharePoint Lists · GitHub Copilot (AI-assisted development)
Website for my father's loft drywall company: professional content, a gallery, and a GDPR-compliant contact form with reCAPTCHA, legal notice and privacy policy.
React · Bootstrap / React-Bootstrap · AOS (scroll animations) · Formspree + reCAPTCHA — view it →
Website for a console, VR and simulator rental business: two entry points (event / home), an interactive zone planner, content editable without code.
Next.js · TypeScript · Tailwind CSS · Cloudflare Workers — view it →
I build internal tools and process automation, working with AI agents — measurably, and with a human always making the call on what goes live.
bence@portfolio:~$ whoami
Bence Krisztián-Világos — full-stack developer background (JavaScript & C#), building with AI-agent-based workflows.
bence@portfolio:~$ cat job.txt
I'm a logistician at a payments/fintech company; alongside the role I built a Power Apps internal equipment-request system that runs in production, and I keep developing it.
bence@portfolio:~$ cat experience.txt
Previously (in 2025), I worked as a QA / software tester for six months.
bence@portfolio:~$ ls ./side-projects
football match-prediction ML model [research · case study: case]
this portfolio site [live · engine room: ops]
bence@portfolio:~$ cat interests.txt
AI/LLM direction, machine learning — and anything that runs in a terminal.
bence@portfolio:~$
featured case study
How I killed my own backtest
A +6.42% backtest return that came entirely from data leakage. An interactive measurement, the mechanism of the bug, and the rules that have guarded the system since.
+6.42% → +0.68%
read it →
Colleagues used to flag equipment requests in Teams messages, which we then logged one by one in personal spreadsheets — incomplete info, untraceable requests, data scattered everywhere. That got replaced with a Power Apps system where requests (for business units or events), issuing, returns, receiving, and transfers are all tracked digitally, down to serial/inventory number and device configuration, with full traceability. Live in production: the full request-to-issue workflow, a dashboard, and an inventory-audit view; a stock overview and a per-device history view are still in development.
Built a website for my father's company, Vigo Bausystem Kft. (a loft/attic drywall specialist): showcases fire-protection and thermal/sound insulation solutions, vapor-barrier technology, partition walls, and suspended ceilings with professional content, a gallery, and a GDPR-compliant contact form (Formspree, reCAPTCHA-protected) — plus imprint and privacy-policy pages.
New website for a business that rents out consoles, VR and simulators for events and for home use. A fully static Next.js site with two entry points (event / home rental), an interactive zone planner that shows the number of stations, simultaneous players and staff for each package, and a step-by-step quote request page. Texts and prices live in JSON files and can be edited without code; hosting is on the Cloudflare Workers free plan.
An experimental, research-focused machine learning system for predicting football match outcomes (home/draw/away) and Over/Under 2.5 goals across three leagues (Premier League, Bundesliga, La Liga) — not a tipping service. The XGBoost-based model is complemented by a Poisson-distribution statistical goal model, local LLM-based (Ollama) probability correction, and a rule-based knowledge layer. Validation uses walk-forward testing on a held-out dataset, evaluated via closing line value (CLV) — the automated agent never touches production.
This very site — built on Next.js and Tailwind, with a terminal-inspired look. Keeps growing as new projects and skills get added.
{ "languages": ["JavaScript", "TypeScript", "C#", "Python"], "frontend": ["React", "Next.js", "Tailwind CSS"], "backend": [".NET", "Node.js"], "data_ml": ["pandas", "scikit-learn", "XGBoost", "SciPy", "DuckDB"], "ai_llm": ["Ollama / local LLMs", "prompt engineering"], "automation": ["Power Apps (Canvas app)", "Power Automate", "SharePoint Lists"], "tools": ["Git", "CI/CD (Vercel)", "scheduled jobs (cron)"] }