Abdelkrim
El Ayachi

SOC analysts drown in Wazuh alerts. Triaging each one means pivoting by hand across alert history, past investigations, and asset records — slow, repetitive, and easy to get wrong under volume.

Give an autonomous agent the same read-only tools an analyst uses and let it run the triage loop: read the alert, decide what context it needs, call a tool, read the result, repeat — until it reaches a verdict, with every step recorded for a human to audit.

A Python/FastAPI backend polls the Wazuh Indexer on an interval; a policy engine decides which alerts open an investigation; each one is handed to an agentic tool-calling loop over MySQL and OpenSearch. The LLM layer is provider-agnostic — any OpenAI-compatible endpoint or a local Ollama model, switched with one env variable. A Vue 3 + Pinia dashboard surfaces verdicts, evidence, and the full action timeline.

The AI is deliberately walled off from all user-identifying data — every tool query excludes it — so analyst attribution can't bias a verdict. Hard budgets on steps, tool-result size, and total conversation cap cost and stop indecisive loops; any failure still records an ERROR verdict, so nothing hangs at IN_PROGRESS.

Next: authentication and a way to cancel a running investigation. After that, extend the tool architecture additively — Suricata, CloudTrail, threat-intel enrichment — as new tool classes rather than a redesign of the loop.

Typing practice tools are boring and solo. You grind through exercises alone, with no reason to come back. There's no energy, no stakes.

What if typing practice felt like a race? Real-time, multiplayer, with live scores updating as you type. Make it competitive and people will actually practice.

A real-time multiplayer typing app with live rooms, instant score updates via WebSockets, and dynamic content pulled from the Wikipedia API so you're never typing the same thing twice. Redis handles caching, JWT handles auth.

Chose Redis over persistent storage for session data — speed mattered more than durability here. Used Socket.io instead of raw WebSockets for automatic reconnection and fallback support. Worth the extra bundle size.

Better error handling on socket disconnects — users sometimes get stuck in ghost rooms. I'd also add persistent leaderboards instead of session-only rankings, and rethink the room cleanup logic.

Every YouTube downloader I tried was either bloated with ads, sketchy, or gave you low-quality output. I just wanted the best quality without the noise.

YouTube serves audio and video as separate streams at the highest quality. If I download them separately and merge with FFmpeg, I get the best possible result with a simple pipeline.

A lightweight desktop app that downloads the highest quality audio and video streams separately, merges them with FFmpeg, and gives you a clean output. No accounts, no ads, no bloat.

Used pytubefix over the original pytube — more reliable but less documented. Went with a desktop UI (pywebview) instead of a web app because this is a personal tool, not a product. Simplicity won.

Add batch downloading, a more accurate progress bar (FFmpeg doesn't make this easy), and maybe a queue system for large downloads.

Vectis is INVOLYS's ERP suite — a large, module-heavy business platform where a regression in one module can quietly break workflows in another.

A broad functional surface, frequent releases, and deep data dependencies between modules — manual regression alone couldn't keep pace with the release cadence.

Ran functional and regression testing for assigned modules: writing and executing test cases, reproducing and documenting defects in Jira, and following them through the defect lifecycle with the development team.

Contributed to a Selenium regression suite for the highest-risk flows, which reduced repetitive manual regression work and freed time for exploratory testing.

A faster, more trustworthy regression signal on every release, and earlier defect detection — before changes reached UAT.

Praxis is another INVOLYS business application I test alongside Vectis, with its own release stream and user workflows.

Validating new features and change requests against existing behaviour without a full automated safety net, under tight release windows.

Ran functional testing on new features and change requests, executed regression passes on each build, and validated API behaviour with Postman where the UI didn't tell the whole story.

Risk-based prioritisation — depth where regression is most likely or most costly — paired with clear, reproducible defect reports to keep the feedback loop with developers short.

Consistent release validation and fewer defects escaping to production.