Artur SolodovnikovAI-first developerShort CV
A research lab staffed by AI agents.
Product, backend, frontend, smart contract, infrastructure · since February 2026
82 agents signed up on their own. Every light here is an agent.
An agent reads skill.md, gets an API key and picks up research tasks by itself.
How the pipeline works
Every dot on the floor is a fact that passed an independent audit.
The token and its fees paid agents real USDC for audited work. as of
Python, FastAPI, PostgreSQL + pgvector · Next.js 16, React 19 · Solana: USDC payouts, staking — a Rust program on mainnet · 481 commits · ≈119,000 lines · 926 tests
The Observatory on open-solve.com: domains, questions and tasks as one graph, and bottom left the feed of agents approving facts
Open open-solve.comThe protocol on GitHubThe smart contract on Solana
A neuromorphic trading agent.
The frog can’t see the flies — it feels the ripples.
Product, code
A spiking neural network modelled on the nervous system of the clawed frog Xenopus laevis: 333 neurons, 2,938 synapses.
New trading pairs are flies on the water. The frog decides whether to swallow, spit out or digest. It learns by itself, through dopamine, from the outcome of each trade. No one steers it by hand.
Inside the brain
Node.js, WebSocket, three.js · the spiking model is my own code
Digital Jungian analysis. An AI analyst that reflects rather than comforts: it remembers your dreams and notices where you sidestep the question.
A Telegram bot, a website and a Mini App. A new version is in progress.
Product, design, code TypeScript, Fastify, grammY, PostgreSQL, SvelteKit, OGL
The heart of the new version isn’t a screen, it’s memory: what the analyst knows about a person, and where it learned it.
Dreams
Kept in the dreamer’s own words, not the model’s retelling. Dream-dictionary readings are thrown out; dreams are linked into series by shared images — Jung read dreams in series.
Only their words
Every observation about a person rests on a verbatim quote. If the quote isn’t in the conversation, the observation is dropped: the analyst can’t attribute to someone what they never said.
Patterns of character
Named in the person’s own words, never as diagnoses. They grow stronger with confirmations across different conversations and fade over time; pushback under pressure doesn’t count as confirmation. Code does the counting — the model only reads and classifies.
Safety over depth
In a crisis, work with the Shadow switches off entirely — nothing the analyst is otherwise allowed to do overrides that.
The path: Persona, Shadow, Anima and Animus, Sage, Self. The bot, the website and the Mini App share one memory; voice messages are transcribed.
The landing page was built with the Gauntlet Loop: one agent builds, three critics compare the first screen blind with Awwwards-level references. Lose, and it’s another round. The critics wrote in Russian; their words are translated.
I work with AI agents the way you’d run a small team: I set the task, split it into parts and check each one. I read the code and the logs myself.
That’s how the ShadowJung landing page was built.
Today I work in Claude Code; for a long time before that, in Cursor. I’ve tried Codex. I use ready-made instruction sets for agents and give them separate roles: one builds, one checks, one gathers material. I read the code and the logs myself: if an agent takes a wrong turn, I’ll see it and sort it out.
No fixed process: I look, try and compare. I want a thing to look like nothing else and still be simple and easy to use. Awwwards-level references and blind critics decide which version is stronger.
Only what I have already used in my own projects.
Work I have already done in my own products.
Routine that runs by itself, with no manual work: payouts in cycles, contract checks, scheduled messages. I’ve also worked with n8n.
Already built: OpenSolve, XENEURO, ShadowJung
Scientific databases (PubMed, ArXiv, Crossref), market data (DexScreener, GeckoTerminal), sign-in with Clerk and Telegram, crypto wallets.
Already built: OpenSolve, XENEURO, ShadowJung
Conversation, memory of the person, onboarding, and messages the bot sends first.
Already built: ShadowJung
Agents that sign up on their own, take tasks and follow a protocol.
Already built: OpenSolve
FastAPI and PostgreSQL, auth, database migrations. It’s live and agents use it.
Already built: OpenSolve
A screen that shows what is going on: live figures, a graph of tasks, a feed of events.
Already built: OpenSolve
Code that ships with tests, so changes don’t break what already works: 926 tests in OpenSolve, 32 for the smart contract, plus an audit report.
Already built: OpenSolve
A first screen that beat two of three Awwwards-level references in a blind comparison.
Already built: ShadowJung
Before code, I worked with people: support at Sber and Yandex, key accounts at PIK. So I know what a product looks like to the person using it.
Artur Solodovnikov · AI-first developer
Artur Solodovnikov
AI-first developer · Tula, Russia · remote only
I build products with AI agents — from idea to production.
Telegram @riviaConfigsolodovikov.aa@yandex.rugithub.com/cosArthro
A research lab staffed by AI agents: agents sign up through the API on their own and take research tasks; a fact is accepted only when independent auditors agree.
Product, backend, frontend, smart contract, infrastructure · since February 2026
Digital Jungian analysis: an AI analyst that reflects rather than comforts. A Telegram bot, a website and a Mini App; v2 is in progress.
Product, design, code
A neuromorphic trading agent: a spiking neural network modelled on the nervous system of the clawed frog, 333 neurons, 2,938 synapses. Live DexScreener and GeckoTerminal data, simulation mode.
Product, code
I work with AI agents the way you’d run a small team: I write their rules (CLAUDE.md), split the work into parts with a clear definition of done, have a separate critic agent compare the result blind against a real reference, accept code only with tests and keep a decision log. I read the code and the logs myself and understand how it all works. My main tool is Claude Code; for a long time before that, Cursor.
Education: Tula State University, incomplete degree · Language: Russian — native
