ORPHEUS
A 139.7-second reel of Bulgarian history, rendered from code, where each picture morphs into the next. 77 of 77 shot starts land within one frame of their downbeat.
Drawn in typea (2,3) torus knot
GeorgiDimitrov
- Head of AI Engineering, DigiSol
- Founder, ValoxVSL
I direct fleets of AI agents that build and verify software, and I design the harness they work inside.
- BSc Pharmacology, First Class
- Varna, Bulgaria
Agents write most of my code.
I specify the work, direct the fleet and read the evidence before anything ships. Agents handle the language and the tools. Every number comes from deterministic code, and a person approves each decision that has consequences.
A replay match report, recreated with synthetic values.
Customs and Compliance Agents
AI agents that prepare customs and compliance filings, across three families of regulated documents. The model does no arithmetic: each figure comes from a deterministic engine, and an engine earned trust only after it reproduced the client's own past filings digit for digit.
Each department has one plain LangGraph agent and no supervisor above it. A change to the data parks on a signed approval card until a person accepts it, and each write lands in an append-only audit table.
- 135 / 135amounts of a closed permit reproduced by the incentive-closure engine
- 1,185 / 1,185cells of the client's own consumption matrix recomputed identically from raw parameters
- 166 / 166capacity-report rows recomputed exactly and matched to the signed final reports
- 24 / 24lines of an approved filing reproduced, kept as a permanent regression test
The tool: references go in, scenes with the same faces come out.
ValoxVSL
A script goes in and a finished ad or documentary comes out. A Visual Bible pins the characters, places and continuity rules before any pixel exists, and a person approves it and the reference images. Narrow agents then handle each scene, and an FFmpeg editor cuts on the voice.
Video generation submits each scene to the provider queue first and polls afterwards, behind key rotation and a cost table for each call. An MCP layer lets Claude act as the producer, and an agent has driven a 67-scene, 179-second ad through it end to end.
The studio has delivered ads for clients, and the tool runs in production behind a sign-in. My brother works on the studio with me.
- 83MCP tools that let an agent run a job end to end
- 38agent configurations in the pipeline
- 14reference images per generated shot, so a face stays the same face
My own avatar on the office floor. The projects and names are demo data.
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 manager agent runs each floor through 20 custom MCP tools: it hires workers, reads their screens and merges only pull requests whose checks are green.
Around the scene I built the harness: workers isolated by the operating system, guests denied by default, and runners that keep logins on your own computer.
- 20custom MCP tools the floor manager runs the floor with
- 32unprivileged pool users, one per AI account, in 16 floor groups
- 20.1 to 6.0 msmedian frame time with 12 streaming workers
- 9independent verification rounds before the PC runner shipped
When an agent reaches for a dangerous command, the hook refuses and says what to do.
Meta Harness
Rules, guards, verification checks and coordination, installed on a machine with one command and no credentials. When an agent fails, I change the harness so the mistake does not repeat. Given the choice between more instruction text and a check that fails, I add the check.
The one custom subagent is a read-only verifier with a fresh context. It answers PASS, REVISE or BLOCK, keys each finding to a file and line, and NOT RUN is a legal answer.
- 12always-on platform rules installed on each machine
- 9workflow skills, each with at least one positive trigger case
- 90table cases in the command guard's selftest, 57 block and 33 allow
- 39trigger eval cases, 6 of them guarding against false triggers
A trick under an All-Trumps contract, with the bidding and score in the corners.
PlayBelote
A pure TypeScript rules engine shared by client and server, Monte Carlo bots that sample the hidden hands, Swiss and knockout tournaments, and BGN, a replay notation modelled on chess PGN.
A second table renderer was built in long unattended agent runs. It had to match the original frame by frame on seeded games, judged by a panel of vision models, and Phaser stays the default until I switch it.
- 60simulated deals behind each card a bot plays, 40 behind each bid
- 55%simulated win rate a bot needs before it opens the bidding
- 13,123bot games across 20 championships of 128 teams
- 10/10seeded game pairings at visual parity between the two renderers
An order book, recorded once a second.
Polymarket research desk
Thousands of configurations and 15 strategy families, tested against live order books under rules written down before any scoring. Each candidate is frozen and hash-stamped first, then scored once.
The money loop is a deterministic daemon with no language model in it. Going live takes three things at once: live mode, an ARMED flag and no HALT file. The survivors run behind kill switches, and the results stay private.
- 15strategy families tested against live order books
- 3,072configurations searched, each candidate frozen before it was scored
- 4parallel research lanes, each checking the others' work
- 0language models in the money loop
Jeisan, as he first appeared in a ValoxVSL ad. AI generated.
Jeisan
He began as a character in a ValoxVSL watch ad, a tough-guy action hero. The name now belongs to my agents: the always-on agent on my server, reachable over Telegram, and the git author Jeisan Steisan on their commits. He also manages the trading pods through a bounded control plane that ships disarmed.
Codex moved him from OpenClaw to NousResearch's Hermes Agent, with a preview, a restore point and a parity audit against the old workspace.
- 52skill packages carried over in the move to Hermes Agent
- 4schedules restored after the move
- 15 / 15same-host isolation checks passed before the worker plane went live
Workshop
Smaller pieces: shipped, paused or unreleased.
PlayBelote v2
A clean-room 3D rebuild with a tavern cast at the table, where bots search up to 48 sampled worlds per decision and solve the last two cards exactly. Unreleased.
Also on the bench: Sesamebot, a Discord bridge that makes a voice channel sound like one caller; GeneScope; Hayzunxn, a joke language kept like software; Mafia with AI; Tuzzo Music; programmatic music in odd meters; and Sapverse, a 24-chapter novella written with early AI models, my first AI project.
The real site, whole, in a browser window.
Biogard
The first project I started. My oldest saved chat is about a React app for it that would not start. It moved to Next.js on Netlify, and later I rebuilt it from scratch with the same URLs, hand-drawn illustrations and a self-hosted, isolated stack.
The site hides a bug game. So does this page.
- 100mobile PageSpeed score in all four categories (home page)
- 38hand-drawn SVGs: 27 pests, 9 places, 2 scenes
- 122pages crawled with 0 problems
Write to me
datuzzo7@gmail.comValoxVSL work: sales@valoxvsl.com
Next: Valox Cinema, a fully AI feature film. No film exists yet, and two cheap tests come before any production.