Veneto, Italy

Gabriele Masiero

Final‑year Computer Science Student

Software & AI Engineer • Multi‑Agent RL

Thesis: Multi‑Agent RL (MAPPO/MADDPG) BenchMARL • PettingZoo • VMAS Open to roles in Veneto

I build pragmatic software and love turning research into impact. Currently finishing my thesis in Multi‑Agent Reinforcement Learning and actively looking for roles in Veneto.

Download Resume View Projects Contact Me

Portfolio • dixieoff.me• GitHub• LinkedIn• Dev Checklist

About Me

I’m a builder at heart. From web platforms to research tooling, I value clean code, fast feedback loops, and delivering things people actually use. Recently I’ve worked with BenchMARL, PettingZoo, and VMAS while exploring algorithms like MAPPO, MADDPG, and MASAC.

My next step: contribute to a product team where I can own features end‑to‑end, keep learning, and bring applied AI to real users.

What I'm looking for

• Backend / Full‑Stack (TS/Go/Python)
• AI/ML Engineer (applied)
• Teams in Veneto (Padova/Verona/etc.)

KLens Core Developer

Core developer and maintainer of KLens — NestJS + MongoDB document intelligence platform for Italian legal and real estate contracts. Built production-ready AI extraction with 28 document types, per-type key-value extraction, and hardened chat interfaces (100% jailbreak resistance in stress tests).

Tech Stack

NestJS backend, MongoDB, custom AI extraction pipelines. Supports multiple tenants. Handles complex Italian contract parsing with ML-powered field extraction.

Production Ready

Stress-tested May 8: ALL jailbreak queries blocked. Multi-tenant architecture with isolated data, RESTful APIs, and real-time document processing. Currently serving production workloads for Italian legal tech.

lele-harness

Most AI agent frameworks are black boxes — you configure them and hope. lele-harness runs on the opposite philosophy: a CLI-first agent harness small enough to read end to end.

A 535-line ReAct loop you can read in an afternoon. Core actions are native JSON tools — exactly the shape models are trained to call. Everything else is just an executable on your PATH, in any language. No framework to fight, no SDK, no platform lock-in — an agent that works with you, not behind your back.

You Can Read the Whole Thing

The loop is a single 535-line file — no agent framework underneath, no hidden control flow. Tools are native JSON — read_file, write_file, edit_file, bash — the exact function-calling format models are trained on. What you see is what the model gets. No surprises.

You Own the Stack

Any OpenAI-compatible provider — OpenRouter, a local vLLM, whatever you run. Skills are executables on PATH in any language; sub-agents are just markdown files in a folder, orchestrated in parallel from a small script. No SDK. No lock-in. No “works on my machine.”

You Talk While It Works

Type while the agent runs — your message drops into the live turn. Auto-compaction kicks in at 80% of the model’s real context window, and post-turn compile checks feed errors straight back to the model. Built for people who code with AI, not people who babysit it.

Beyond the Code

AI Tech Wizard

Deep in AI tech: stays current on every breakthrough, builds BTC debate engines, synthetic datasets, fine-tunes with axolotl/unsloth, runs local inference. Not a vibecoder.

Self-Hosting Pro

Homelab with Navidrome, Tailscale, pet speech bubble setup (F8→webhook→MQTT→WS). Plus tmux continuity daemon for persistent memory systems.

Aesthetic Purist

Demands dark themes, proper typography, pixel text with zero anti-aliasing, self-hosted Press Start 2P font. High standards, no compromise on visuals.