Georgi DimitrovdaTuzzoContact

daTuzzo, Varna, Bulgaria

Georgi Dimitrov

Pharmacologist by training, Head of AI Engineering at DigiSol. I run fleets of AI agents that write software, and I build the checks that decide what ships.

Two ways in. The dot on the i is a whole world.

Scrolling works too, and the choice is remembered.

Scroll, or press play. The dot on the i is a whole world.

Inside the dot

The system I run

Each dot is a project, 40 in all, and each line is a real link between two of them. The ring is the Meta Harness, under every repo: the rules, guardrails and checks coding agents work inside. When an agent fails, I fix the harness so the mistake does not repeat.

Terminal output of the harness's PreToolUse hooks. Four simulated agent actions are each refused with exit code 2, an orange BLOCKED line stating the reason and a line giving the remedy: a blind git commit with -a, a recursive delete of the home directory, reading a .env file and writing to a .env file. A scoped git add is allowed with exit code 0, and the last line reports the guard self test passing.The harness at work: its hooks refuse a blind commit, a recursive delete of home and reading a .env file.

System / ValoxVSL

ValoxVSL

My AI video production company, and the pipeline it runs on: a script goes in, a finished film comes out. Narrow agents work each phase and a deterministic FFmpeg editor makes the cut. The hard part is continuity, so a Visual Bible fixes every character before any pixels exist.

ValoxVSL, AI video ads

ValoxVSL / one scene

One face, six shots

An AI soldier, Ivan The Defender in the tool, in six shots of a Doiran documentary ad, all generated from one character reference. Each image call gets up to 14 references, so the lean face, the pale eyes and the mud hold from shot to shot.

System / Valox Office

Valox Office

A 3D office where Claude Code, Codex and OpenCode workers sit at desks, each in its own git worktree, and teammates walk the same floors. It is a fork of the MIT project AgentSystemLabs/agent-office; my fork adds the floor managers, per-account worker isolation and runners on people's own PCs.

Valox Office / Office 1

A floor is a project

I walk each floor as my own avatar, among the workers and the boards. A floor manager, itself a Claude Code or Codex worker, hires and directs the others through 20 MCP tools, and it merges only pull requests whose checks pass.

Floor / the whiteboard

The floor's whiteboard

The first tab is a shared drawing board; agents add their own tabs through the whiteboard tool, as Mermaid, drawings or HTML. Agent HTML renders in a sandbox with no script, so nothing an agent draws can reach into the office.

Office 1 / the manager's desk

The floor manager's desk

The manager works at a laptop like every worker, each in its own git worktree. On Linux each AI account runs as its own unprivileged user, one of a pool of 32, and that isolation was denial-tested on a throwaway server before it went live.

Desk / the manager's screen

The manager's screen

A shared terminal that teammates can watch live. The manager lists issues, reads a worker's screen with read_worker_screen and hires through its own tools; workers answer through report_to_manager.

Screen / one tool call

One tool call

An MCP tools/call request with one argument, delivered to the floor manager's inbox. In the first end-to-end run, a manager hired a researcher to say pong, asked the humans before hiring a second worker, then hired that one to say ping.

System / PlayBelote

PlayBelote

Live multiplayer Bulgarian belote. One deterministic TypeScript rules engine runs in the browser and on the Colyseus game server, so both sides apply the same rules, and a player who keeps sending illegal moves is disconnected.

PlayBelote, play it live

PlayBelote / one hand

A hand

Eight cards from a 32-card deck. The server records every game in BGN, a replay format modelled on chess PGN, and the viewer replays it from any seat.

Hand / one card

One card

The bot at South has to choose a card while 21 cards stay hidden from it. Bot types are ranked by tournament: 20 championships of 128 teams, an ELO for each type.

Card / the search inside

The search inside the decision

The Monte Carlo player deals the hidden cards at random, plays each legal card through 60 simulated games, and keeps the card with the best average. Bids get 40 simulations.

System / trading desk

Trading research desk

A lab for Polymarket's 5 and 15 minute crypto up or down markets, running live. A daemon ticks once a second: it records books and trades, quotes in shadow mode and settles each window at resolution. No language model sits in the money loop.

Trading desk / one order book

One order book

A shadow bid counts as filled only when the trade tape prints through it. Thousands of configurations and 15 strategy families have met live order books this way, under rules frozen before scoring. The survivors run behind kill switches, and the results stay private.

Trading desk / operator overlay

The operator overlay

My Chromium extension draws over the exchange page: it follows each live window, shows the pod's state and edits only settings the server whitelists. HALT has its own button, and going live takes the exact typed phrase ARM and the pod's name.

Order book / kill switch

The kill switch

A live order needs three locks open at once: mode set to live, an ARMED file a person creates, and no HALT file. Touching HALT cancels every quote, and a daily loss stop writes it by itself.

System / more worlds

More worlds in the dot

  • Customs and Compliance Agents for a large industrial manufacturer: agents handle the language, deterministic engines produce each figure, a person approves each data change.
  • A bald, square-jawed man in a black leather jacket leans against the wheel arch of a dark car on a wet night street, an orange and black watch on his wrist, looking into the camera with a half smile. Behind him, out of focus, a man in a navy suit and a woman in a black dress look at a phone. Every person in the frame is AI generated.Jeisan, the name my working agents carry: an always-on server agent, a git author and the manager of the trading pods.
  • Valox Cinema, a fully AI feature film planned on the ValoxVSL pipeline, starting with two tests that could stop it.
  • ORPHEUS frame: the headline Krum raised a toast in red and white beside an engraved skull cup with a silver rim and a red liquid surface, red rays behind it on a dark ground.ORPHEUS, a code-rendered reel of Bulgarian history in which every picture morphs into the next.
  • Home page of the Biogard pest-control site: a dark hero with a Bulgarian headline about a pest-control firm in Ruse and the region, a white line drawing of a house in cross-section inside a dashed red outline, ants walking toward it and a mosquito overhead, and a request-inspection button. Phone numbers and the e-mail address are masked.Biogard, my first project: a regional pest-control company's site, rebuilt self-hosted and live.

Biogard, the live pest-control site

Pulling back out

Georgi Dimitrov

Varna, 43.2° N, 27.9° E

Where the dot sits

I live in Varna, on the Black Sea, and grew up in Ruse, on the Danube. Pharmacology took me to Portsmouth first. Write to me about agents and the harnesses that keep them honest.

daTuzzo, Varna, Bulgaria

Georgı Dımıtrov

Pharmacologist by training, Head of AI Engineering at DigiSol. I run fleets of AI agents that write software, and I build the checks that decide what ships.

Third-person view of the author's own 3D avatar standing on the open office floor: a man with dark curly hair and a short beard in a navy blazer with red embroidered cuffs, a black shirt and olive trousers. Behind him are desks with coding agent workers under name labels, the issues board, the task queue, a pull requests board, a whiteboard showing a checkout funnel chart, a potted plant and a gong.
Walking the floor as my own avatar, in my fork of the MIT-licensed AgentSystemLabs/agent-office. Every worker is a coding agent; the project and names are demo data.

01 / The rules, guards and checks my coding agents work inside, across Claude Code and Codex

The system I run

12
always-on platform rules installed on each machine
9
workflow skills, each with at least one positive trigger case
8
projects with their own project-layer rule sets

Each dot is a project, 40 in all, and each line is a real link between two of them. The ring is the Meta Harness, under every repo: the rules, guardrails and checks coding agents work inside. When an agent fails, I fix the harness so the mistake does not repeat.

Terminal output of the harness's PreToolUse hooks. Four simulated agent actions are each refused with exit code 2, an orange BLOCKED line stating the reason and a line giving the remedy: a blind git commit with -a, a recursive delete of the home directory, reading a .env file and writing to a .env file. A scoped git add is allowed with exit code 0, and the last line reports the guard self test passing.
When an agent reaches for a dangerous command, the hook refuses and says what to do instead.
The full Meta Harness story

02 / My AI video-ad studio and the multi-agent pipeline it was built on

ValoxVSL

Six AI-generated frames of one fictional Bulgarian infantry soldier, called Ivan The Defender in the tool, from six different shots of a Doiran documentary ad: crouched in a trench doorway in front of a fire, looking up in a dim bunker, lying wounded on his back, charging with a rifle, aiming at the camera with wet hair, and wearing a French style helmet in grey light. The lean face, pale eyes and mud stay the same in every frame.
One AI character in six scenes, all generated from the same character reference.
83
MCP tools that let an agent run a job end to end
14
reference images per shot, so a face, a product label or a room stays fixed
67 scenes
in a 179-second ad that an agent produced end to end through the MCP

My AI video production company, and the pipeline it runs on: a script goes in, a finished film comes out. Narrow agents work each phase and a deterministic FFmpeg editor makes the cut. The hard part is continuity, so a Visual Bible fixes every character before any pixels exist.

An AI soldier, Ivan The Defender in the tool, in six shots of a Doiran documentary ad, all generated from one character reference. Each image call gets up to 14 references, so the lean face, the pale eyes and the mud hold from shot to shot.

The Jeisan character in the watch ad, on the street and behind the wheel.
AI-generated depiction of General Vazov, the Bulgarian commander at Doiran in 1918, in three scenes of the same documentary ad: a large close-up in a peaked cap with heavy brows and a thick grey moustache, a standing portrait in a wooden command hut wearing a greatcoat, and a portrait in a trench in a field tunic holding binoculars under a smoky sky. The face is an AI depiction of a historical figure, not a photograph.
General Vazov, an AI depiction, in three scenes with three outfits and three lightings.
The full ValoxVSL story

03 / An open-source 3D agent office, stripped of the slop and rebuilt as a harness for agents

Valox Office

The Valox Office tower seen from outside in a stylised city: a four-storey beige building with balconies and terraces, a dark glass band along the top, a lift housing on the roof and a garage with parked cars at ground level, on a tree-lined street among pastel skyscrapers.
The tower from outside: offices with terraces, the executive floor's glass band and the garage.
20
MCP tools the floor manager drives, plus 5 desk tools every worker gets
32
isolated worker accounts in 16 floor groups, each an unprivileged pool user
9
independent verification rounds before the PC runner shipped

A 3D office where Claude Code, Codex and OpenCode workers sit at desks, each in its own git worktree, and teammates walk the same floors. It is a fork of the MIT project AgentSystemLabs/agent-office; my fork adds the floor managers, per-account worker isolation and runners on people's own PCs.

I walk each floor as my own avatar, among the workers and the boards. A floor manager, itself a Claude Code or Codex worker, hires and directs the others through 20 MCP tools, and it merges only pull requests whose checks pass.

The first tab is a shared drawing board; agents add their own tabs through the whiteboard tool, as Mermaid, drawings or HTML. Agent HTML renders in a sandbox with no script, so nothing an agent draws can reach into the office.

Over the shoulder of the floor manager, a worker figure with an orange spiked head seated at the boss's desk in front of a laptop that shows terminal text. Through the glass wall behind it are the open floor, a Task queue board, a Pull Requests board, a lounge sofa and a lift marked Storefront, and two empty chairs face the desk.
The floor manager at work, with the floor beyond the glass.
First-person view of a whiteboard in the office showing a worker's Checkout funnel page: horizontal green bars for product page, cart, checkout and paid, with three figures below them. An empty task queue board, a gong and a media board stand to either side, and a hint at the bottom offers to draw together, live.
A worker's HTML page on the whiteboard, shown in a sandboxed frame.
The full Valox Office story

04 / Live multiplayer belote: a deterministic engine, Monte Carlo bots and a replay format

PlayBelote

A PlayBelote table mid-hand under an All-Trumps contract. Three bots hold fanned hands at the top, left and right, the player's cards are fanned along the bottom, four cards lie in the centre, and the bidding history, score and announcements sit in the corners.
A trick under an All-Trumps contract, with the bidding history and score in the corners.
13,123
bot games across 20 championships of 128 teams
12
bot types rated by ELO in the 20-championship run
1.4
current version of BGN, the replay format modelled on chess PGN; the viewer switches perspective between seats

Live multiplayer Bulgarian belote. One deterministic TypeScript rules engine runs in the browser and on the Colyseus game server, so both sides apply the same rules, and a player who keeps sending illegal moves is disconnected.

The Monte Carlo player deals the hidden cards at random, plays each legal card through 60 simulated games, and keeps the card with the best average. Bids get 40 simulations.

PlayBelote, play live Bulgarian belote in the browserThe full PlayBelote story

05 / A fail-closed trading desk where every strategy meets frozen rules and live order books

Trading research desk

mode = liveARMED fileno HALT filepodexchange
15
strategy families tested against live order books by parallel research lanes
3,072
configurations searched in one sweep, each candidate frozen before it was scored
0
language models in the money loop

A lab for Polymarket's 5 and 15 minute crypto up or down markets, running live. A daemon ticks once a second: it records books and trades, quotes in shadow mode and settles each window at resolution. No language model sits in the money loop.

A shadow bid counts as filled only when the trade tape prints through it. Thousands of configurations and 15 strategy families have met live order books this way, under rules frozen before scoring. The survivors run behind kill switches, and the results stay private.

A live order needs three locks open at once: mode set to live, an ARMED file a person creates, and no HALT file. Touching HALT cancels every quote, and a daily loss stop writes it by itself.

The full Polymarket research desk story

06 / AI agents that prepare customs and compliance filings, with every number checked by code

Customs and Compliance Agents

three familiesenginepast filings, matcheda person approves
135 / 135
amounts of a closed permit reproduced by the incentive-closure engine
1,185 / 1,185
cells of the client's own consumption matrix recomputed identically from raw parameters
166 / 166
capacity-report rows recomputed exactly and matched to the signed final reports

AI agents in production that prepare the customs and regulatory compliance filings of a large industrial manufacturer, across three families of regulated documents. Agents handle the language and the tools, deterministic engines produce each figure and must reproduce the client's own past filings exactly, and a person approves each data change.

The full Customs and Compliance Agents story

07 / The name my working agents carry: an always-on server agent, a git author, a pod manager

Jeisan

A bald, square-jawed man in a black leather jacket leans against the wheel arch of a dark car on a wet night street, an orange and black watch on his wrist, looking into the camera with a half smile. Behind him, out of focus, a man in a navy suit and a woman in a black dress look at a phone. Every person in the frame is AI generated.
Jeisan, as he first appeared in a ValoxVSL ad.
5
least-privilege hops between Jeisan and a trading pod, ending at a strict server-side parser
52
skill packages carried over in the move to Hermes Agent
15 / 15
same-host isolation checks passed before the worker plane went live again

Jeisan is the name my working agents carry. It began as a tough-guy action hero in one of the ValoxVSL ads, and today it is an always-on agent on a small Linux server, moved from OpenClaw to NousResearch's Hermes Agent, the git author Jeisan Steisan, and the agent that manages my trading pods through a bounded control plane.

The full Jeisan story

08 / four more

More work

test 1test 2the film, not shot yet

Valox Cinema

Making a fully AI feature film, starting with the two tests that could stop it

41
decisions in the architecture's table, each with its rejected option and a reopen condition
7
generation modalities behind one adapter interface, each model a typed row whose price and verified capabilities are data
The full Valox Cinema story
ORPHEUS frame: the headline Krum raised a toast in red and white beside an engraved skull cup with a silver rim and a red liquid surface, red rays behind it on a dark ground.

ORPHEUS

A code-rendered reel of Bulgarian history in which every picture morphs into the next

139.7 s
runtime, no voice-over
78
shots, defined in whole bars on the score's downbeats
The full ORPHEUS story
Home page of the Biogard pest-control site: a dark hero with a Bulgarian headline about a pest-control firm in Ruse and the region, a white line drawing of a house in cross-section inside a dashed red outline, ants walking toward it and a mosquito overhead, and a request-inspection button. Phone numbers and the e-mail address are masked.

Biogard

My first project: a pest-control company's site, rebuilt self-hosted

100
mobile PageSpeed score in all four categories (home page)
1.5 s
mobile LCP, down from 2.7 s
Biogard, the live pest-control websiteThe full Biogard story
re-infected after each rebuildoutboundhardening playbook

The server compromise

A framework RCE, a botnet, three failed rebuilds and the outbound connection that found it

3
clean-ups and rebuilds re-infected before the root cause
4 h
from first deploy to root cause
The full The server compromise story

29 more, smaller

Archive

  • Sixteen tiles forming eight before and after pairs from the same camera positions on a 3D map of a city square: an overview of the district, a tall column monument, a fountain view, a court house beside a fountain, an arcaded building behind trees, an opera house with red walls, a municipal building and a shopping plaza. The after tiles have softer light, paler paving and more detail.Ruse town square mapA real town square rebuilt as a game map by a lane-based agent team
  • Programmatic videosA studio where agents render videos from code, with a fact sheet for each video
  • Spectrogram of the jazz piece Yogurt Culture Shock in amber on near-black, 3 minutes 21 seconds wide, with thin vertical lines and labels marking sections such as Petri Dish, Culture Shock, Lydian Loiter, Bacillus Bulgaricus, Sour Milk Sermon and Kaval Epilogue and Button.Programmatic musicTwo tracks composed as code: a JavaScript suite in odd meters and a MIDI piece
  • A hand of Belote in the 3D tavern theme: a white-haired man in a flat cap, a woman with red curly hair in an embroidered white blouse and a man in a dark flat cap sit at a red and white checked table, with a trick of cards on the cloth and the player's three remaining cards at the bottom. Interface text is Bulgarian.PlayBelote v2A clean-room 3D belote with a tavern cast, a jazz salon and a table that lives in a worker
  • Five stacked spectrograms, each a bright pink and orange pattern on a near-black background, showing the first seven and a half seconds of a track. From top to bottom they are labelled source, vocals, other, drums and bass: the full mix, then the four parts the pipeline separated from it. Frequency runs from 64 to above 4,000 hertz.osu! map generatorTurns an audio file into playable osu! maps where each note sits on an audible sound
  • Agent link shortenerA URL shortener for agent replies on Cloudflare Workers, written by my own agent
  • datuzzo.comThis site: the first version, and the rebuild that replaced it
  • Job page of the transcribe service for a demo recording of a weekly sync: status done, a pipeline strip with speech, video and screenshot stages, a detail pane, a log tail and a list of output files such as transcript, meeting record and meeting notes.Transcribe serviceLong recordings in, a diarised transcript and a screen timeline out, with each run costed
  • Tuzzo VoicePush-to-talk voice control of a Windows PC over a realtime speech model
  • Lead scraperFinds Bulgarian advertisers on paid Meta ads, scores them and tracks outreach
  • SesamebotA Discord bridge that makes a whole voice channel sound like one caller to a voice AI
  • VitalsA private platform for personal health data, kept off this site
  • Chess archive pipelineMy chess.com archive, scraped, scored by Stockfish and charted
  • Tabs CleanupA Chrome and Opera extension that sorts tab piles and stashes them locally
  • HayzunxnA joke language kept like software: a TSV lexicon, a versioned spec and 12 lints
  • A two by two grid of four black and white parcel-style labels for cabinet panels. Each has a black header with a part code, large dimensions in millimetres, role and material lines, four squares marked T, B, L and R showing which edges are banded, a small rectangle diagram and a barcode.CAD/CAM bridgeReverse-engineering a furniture maker's 3D scene files into machine-ready cutting programs
  • Valox InvoiceA self-hosted invoicing tool that issues compliant bilingual invoices and fails loudly
  • Open BrainA personal memory store with semantic search that any MCP client can use
  • YouTube script factoryA Claude Code workspace that drafts and quality-checks documentary narration scripts
  • GeneScopeReads a consumer DNA export against ClinVar and PharmGKB, then four agents interpret it
  • Book writerA Bulgarian-first AI assistant for writing and editing books
  • The Mafia with AI lobby: a purple gradient title, a grid of ten AI players each labelled with a different language model, game settings for discussion time, night time and player count, a colour-coded list of roles and a green Start Game button.Mafia with AIA Town of Salem style Mafia game where the players are language models
  • ValoxVSL site and client portalThe bilingual marketing site and client portal for ValoxVSL
  • The Tuzzo Music app icon: a stylised sound waveform with a small play triangle in the middle, inside a round dark metal disc with a teal glow.Tuzzo MusicA hi-res music player for one household, hardened by a verifying agent fleet
  • Voiceline generatorBatch voice lines from an Excel sheet of prompts, with a Flask UI and a voice library
  • Coach suiteA Next.js dashboard for scouting players and teams from a community game API
  • Marvel Rivals toolsSmall tools for a competitive game group: data pull, analysis, voicelines, teams
  • SapverseA 24-chapter novella written with early AI models, my first AI project
  • Helium hotspot networkA LoRaWAN hotspot network that I deployed and then wound down