Skip to the plain page
daTuzzo

RoofVarna, Bulgaria21:50

Plain page

Head of AI Engineering at DigiSol. Founder of ValoxVSL.

Georgi Dimitrov

The night shift is how I work. I write the brief, and fleets of AI agents build it and review each other while I sleep. I do not type most of the code.

Deterministic code handles every number. In the morning I read the report and approve what ships. Each floor below is one project.

Scroll, or press play

Floor 5, agent platform

Valox Office

A 3D office in the browser where Claude Code, Codex and OpenCode workers sit at desks, each in its own git worktree. A floor manager agent hires them, reads their screens and merges only green pull requests.

I started from AgentSystemLabs/agent-office (MIT), stripped what I call the slop and put in the tooling that makes it a harness: custom MCP tools, runners on remote and personal computers, OS-level isolation and roles.

MCP tools the floor manager drives
20
median frame, from 20.1 ms
6.0 ms
verification rounds before the PC runner shipped
9

Round five found that an office shell could read the session-signing secret and forge an owner cookie. The runner shipped disabled, and it refuses to run until isolation is on.

Floor 4, video production

ValoxVSL

My AI video-ad studio and the multi-agent pipeline behind it. A Visual Bible fixes every character, product and location before any pixels, and a person approves it.

Each scene then passes a classifier, a producer and a prompt engineer, video goes out to Kling, Seedance, Veo and Sora, and an FFmpeg editor cuts on the voiceover. Over MCP, Claude runs the whole job as producer.

MCP tools, so an agent can run a whole job
83
scenes in an ad an agent made end to end, 179 s
67
reference images per shot, so a face stays a face
14

The first agent-made ad never judged motion, so a scene-aware video QA tool went in: Gemini now watches the generated clips.

ValoxVSL, AI video ads

Floor 3, live game

PlayBelote

A live multiplayer belote game. One deterministic TypeScript rules engine runs on client and server, Colyseus holds the rooms, and bots fill the empty seats.

Each bot deals the unseen cards at random, consistent with what the table has shown, and plays the hand out in simulation before it picks a card. Twenty championships of 128 teams ranked the bot types by ELO.

simulations behind every card a bot plays, 40 per bid
60
simulated win rate a bot needs before it bids
55%
replay notation, modelled on chess PGN
BGN 1.4

The tournament run found a leak: finished all-bot rooms were never disposed, and the server ran out of memory at 1.8 GB after tournament 11. All-bot rooms now disconnect when the game ends.

PlayBelote, live belote online

Floor 2, research desk

Polymarket research desk

A desk for 5 and 15 minute crypto markets, split by authority. The money loop is a deterministic daemon with no language model in it, and the agent control plane fails closed: it holds no trading power by default.

Thousands of configurations and 15 strategy families have met live order books under rules frozen before scoring. The survivors run behind kill switches, and the results stay private.

strategy families against live order books
15
configurations in one search, each frozen before it was scored
3,072
language models in the money loop
0

A live order needs three things at once: live mode in the config, an ARMED file and no HALT file. Touching HALT pulls every quote.

Floor 1, first project

Biogard

The first project I started: the website of a regional pest-control company. I rebuilt it from scratch on Next.js 16 with the same URLs, hand-drawn illustrations and a self-hosted, isolated stack.

One brief from me, a scaffold from the lead agent, then eight agents in parallel on pages, blog, legal and SEO. A crawler written alongside them became the quality gate.

problems in a crawl of 122 pages
0
mobile PageSpeed, all four scores
100
species in the hidden bug game, plus a Cockroach King
13

Containers run read-only with every capability dropped, only Caddy publishes a port, and an egress firewall keeps them off private ranges.

Biogard, pest control in Ruse

Ground floor

The morning report

Every night ends in a document: what merged, what failed, and what needs a decision. Agents handle the language and the tools, deterministic code handles every number, and a person approves every decision with consequences. That person is me.

The whole tower

Every floor in one line: from my laptop on the roof to the table in the lobby.

Morning report

Varna, Bulgaria

Filed by the night shift. Each figure comes from a measurement, a file or a log, and the failures sit next to the wins.

  • 9verification rounds before the PC runner shippedValox Office
  • 67scenes in one ad an agent made end to endValoxVSL
  • 60simulations behind every card a bot playsPlayBelote
  • 3conditions that must all hold before a live orderResearch desk
  • 0problems across 122 crawled pagesBiogard

What went wrong

  • Game server out of memory at 1.8 GB after tournament 11. Fixed: all-bot rooms close when the game ends.
  • An office shell could forge an owner cookie. Round five of nine caught it before the runner shipped.
  • The first agent-made ad never judged motion. A scene-aware video QA tool closed the gap.
  • The old Biogard blog went blank when its database ran out of quota. The rebuild is self-hosted.

Needs a person

datuzzo7@gmail.com

Georgi Dimitrov (daTuzzo). BSc Pharmacology, First Class, University of Portsmouth. Self-taught engineer. Varna, Bulgaria.

The shift is over

Explore another design

Back to the start