Domain 1 / 3

Current, verifiable work that spans ML, backend, DevOps, and product — the range only a generalist can cover, and where I do my own tooling.

NoZip — zero-knowledge encrypted file exchange

A production-grade, zero-knowledge encrypted file-exchange SaaS — now in beta.

Problem

Businesses legally required to exchange client documents and conversations in encrypted form are stuck emailing password-protected zips their clients can't handle.

My Role

Designed and built a zero-knowledge SaaS: client-side AES-256-GCM, per-user X25519 PKI with key rotation, multi-tenant Spaces, Stripe billing, and a layered Supabase schema with RLS — the server never sees plaintext or keys.

Result

Released to beta at nozip.ro, backed by ~34 test files + Playwright E2E. Live proof of modern full-stack engineering plus applied cryptography.

Next.js 16React 19TypeScriptSupabaselibsodiumStripe
Live beta — nozip.roFree access — sign up with codeFREEDEMO

Chisel — AI-assistant scaffolder

A one-shot scaffolder that configures AI coding assistants with project-specific context, rules, and specialized agents.

Generated scaffold

.claude/
├── CLAUDE.md# project overview + pointers to rules, agents, docs
├── rules/
│ └── rules.md# auto-loaded workflow rules & conventions
├── agents/# one file per specialized agent
│ └── *.md
└── skills/
├── dev_mainseat/# driver-mode toggles
└── dev_backseat/
.docs/
├── architecture.md# structure, key systems, data flow
├── todo.md# task tracker
└── todo_done.md# completed-task archive

Problem

AI coding assistants start with zero project context, so you re-explain conventions, architecture, and constraints every single session.

My Role

Designed and built it (open-source, MIT). A single Markdown source of truth generates configs for both Claude Code and GitHub Copilot: it auto-discovers the project from manifests and git history, seeds a living `.docs/` knowledge base, and spins up role-specialized agents (architect, reviewer, security, frontend/backend/embedded) — indexing context on demand instead of dumping the whole codebase.

Result

Public and MIT-licensed — the productized version of the AI-assisted scaffolding practice behind my own projects (NoZip included). Concrete, verifiable proof of the workflow, not just a claim.

MarkdownClaude CodeGitHub CopilotAgent design

Mission Control — an agent-orchestration experiment

A personal sandbox for learning how multi-agent LLM orchestration actually behaves — role-specialized agents that try to build, test, and review software across a range of tech stacks.

Problem

I wanted to understand agentic orchestration hands-on, and off-the-shelf frameworks were too token-hungry to experiment with freely.

My Role

Wiring up my own setup on CrewAI to poke at the mechanics: Architect / Developer / Integrator / QA / Supervisor roles, work decomposed into parallel 'waves', a build/test verify-and-rework loop (cmake/ctest), and routing between local and API models to keep experiments cheap (tracking paid-vs-free tokens).

Result

Very much a work-in-progress experiment, not a product — a hands-on way to learn where agentic pipelines break, how far local models can carry the load, and what verification loops it takes to get usable output.

CrewAIPythonLM StudioGLM APIcmake/ctest

deletarr — self-hosted media tool

A full-stack, containerized tool that keeps a torrent client clean against a Radarr/Sonarr library, with a modern web UI.

Problem

In a homelab *arr stack, torrents linger after their media is removed from the library, silently eating storage — and existing cleanup scripts are unsafe and headless.

My Role

Built a safety-first full-stack app (hardlink detection, dry-run, configurable limits) with a web UI, published as multi-arch Docker images via GitHub Actions CI, deployed on TrueNAS SCALE.

Result

Released and running daily in my own homelab — a containerized, CI/CD product owned end to end, born from a real personal need.

PythonJavaScriptDockerGitHub Actions

Power outage aggregator

Backend that unifies every Romanian provider's planned-outage data so people find out before the lights go out.

Problem

Planned maintenance outages are published — but scattered across per-provider PDFs and websites nobody checks. People simply don't know their power is about to be cut.

My Role

Own the backend: a scheduled polling pipeline that ingests and normalizes every provider in the country — some via structured web APIs, some via published PDF notices parsed into a common schema. A separate frontend (by a collaborator) handles presentation.

Result

Nationwide coverage unified into one clean service — with per-address notifications when an outage is scheduled for you. Frontend integration in active development.

BackendData pipelinePDF parsingScheduling