Artur SolodovnikovAI developerCV

OpenSolve

A research lab staffed by AI agents.

Launched 20 July 2026 · sole author: product, backend, frontend, smart contract, infrastructure

Open open-solve.comThe protocol on GitHubThe smart contract on Solana

82 agents signed up on their own. Each light is one of them.

One page of instructions (skill.md) is all an agent needs: it gets a key and picks up tasks on its own.

How the pipeline works

  1. Decomposition. A big question is broken down into research tasks.
  2. Fact gathering. Agents dig through scientific databases: PubMed, arXiv, OpenAlex. Papers are cut into passages, semantic search (RAG) finds the agent the relevant ones, and it submits its work.
  3. Independent audit. Several auditors check the work, each on their own. No agent is taken at its word: a fact is accepted only when the auditors agree.
  4. Synthesis. A synthesizer agent pulls the results together and looks for what’s missing. That becomes the next question.

The dots on the floor are accepted facts.

82
agents · signed up on their own
848
research tasks
3,096
submissions · 737 passed the audit
2,495
USDC paid to agents to date

figures as of

Agents were paid in USDC for audited work, out of the token’s fees.

Python, FastAPI, SQLAlchemy, Pydantic · PostgreSQL + pgvector · RAG: PyMuPDF, sentence-transformers · Next.js 16, React 19, Tailwind, shadcn/ui, Clerk · Solana: USDC payouts, staking — a Rust program on mainnet · Docker Compose, Caddy · 481 commits · ≈119,000 lines · 926 tests

XENEURO

A neuromorphic trading agent: neurobiology turned into code.

The frog can’t see the flies — it feels the ripples.

xeneuro.org

Inside is a spiking neural network modelled on the nervous system of the clawed frog Xenopus laevis: 333 neurons and 2,938 synapses.

New trading pairs land on the water like flies. The lateral line picks up the ripples, and the frog decides whether to swallow, spit out or digest: buy, sell or take profit. It learns from the outcome of every trade, as dopamine changes the strength of its synapses. For now it trades in simulation, on live market data.

Inside the brain

Model
leaky integrate-and-fire, millisecond time steps
Brain
333 neurons, 2,938 synapses: AMPA/NMDA-like and inhibitory
Learning
dopamine modulation of synapses by the outcome of each trade
Data
live DexScreener and GeckoTerminal, contracts checked with GoPlus
Mode
simulation (paper trading)

Node.js, WebSocket, three.js · the spiking model is my own code

A recording from xeneuro.org: the frog “in vivo”, seen from under the water; rings spread from the flies on the surface.
Touch the live scene ↗: The ShadowJung landing page’s first screen: a night lake painted in gouache, a brass armillary sphere above the water, its light lying on the water as a golden path. (opens in a new tab)

ShadowJung

A Jungian AI analyst. It remembers your dreams, notices when you dodge a question, and doesn’t rush to reassure you.

A Telegram bot, a website and a Mini App. A new version is in progress.

Product, design, code TypeScript, Fastify, grammY, PostgreSQL, SvelteKit, OGL

What’s inside

The new version is built around memory: what the analyst knows about a person, and where it learned it.

  • Dreams

    Recorded word for word, never paraphrased. The analyst throws out dream-dictionary readings and links dreams into series by shared images, the way Jung read them.

  • Only their words

    Any observation about a person has to rest on a verbatim quote. No quote in the conversation, no observation: the analyst can’t put words in anyone’s mouth.

  • Patterns of character

    Named in the person’s own words, never as diagnoses. A pattern grows stronger when it comes up again in different conversations, and fades with time. Agreeing under pressure doesn’t count. Code does the counting; the model only reads and sorts.

  • Safety first

    If someone is in a bad way, Shadow work switches off completely, and no setting can turn it back on.

The stages: Persona, Shadow, Anima and Animus, Sage, Self. The bot, the website and the Mini App share one memory; voice messages are transcribed.

One sentence, four worlds

The landing page’s first screen went through four rounds. One AI agent built it; three critic agents compared it with well-known design sites (Igloo Inc, Astral Frontier and Persepolis Reimagined) without knowing which was which. Every lost comparison came back as a list of notes for the next round. The critics wrote in Russian; their words are translated.

  1. Round one: a vermilion drop cap, a gold mandala, and at its centre a figure on a rock before the moon.
    Round one: lost. A mandala with an illuminated initial. “The central image is a worn-out AI trope.” — blind critic
  2. Round two: a gold sun mandala to the right of the headline on deep blue.
    Round two: lost. A sun mandala in gold leaf. “After the intro, it’s a static poster.” — critic, pair with Igloo Inc
  3. Round three: a gold armillary sphere with a glowing eye at its centre, surrounded by black emptiness.
    Round three: lost. An armillary sphere in WebGL2. “The object hangs in a black void with no space around it.” — critic, pair with Igloo Inc
  4. Round four: a night lake, a brass sphere above the water, its light lying on the water as a golden path.
    Round four: beat two of the three references. A night lake in gouache and a brass sphere above the water. “The sentence literally steps over the object, and the typography carries meaning here rather than labelling.” — critic, pair with Astral Frontier

How I work

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.

  1. Rules. I write the agents a brief (CLAUDE.md): what we’re building, what’s off limits, how to check the work.
  2. Decomposition. I cut a big task into pieces small enough that each one is either done or not.
  3. Blind review. One agent builds; another compares the result with a benchmark without knowing which is which. Falls short? Another round. The method is Matt Shumer’s Gauntlet Loop.
  4. Tests. No tests, no merge. OpenSolve has 926 tests.
  5. Logs. I keep a decision log, so it’s always clear why something was done, and work picks up exactly where it stopped.

That’s how the ShadowJung landing page was built.

Path

  1. Tools

    Claude Code is my main tool; before that, Cursor for a long time. I’ve tried Codex. My editor is VS Code. I’ve also worked with OpenClaw and Hermes Agent. Skills (instruction sets for an agent): I use ready-made ones and write my own. Agents get roles: builder, critic, scout. I write small scripts for my own everyday tasks and use LLMs all the time, not only for work.

  2. Design

    No fixed process. I like things simple and natural, with some character, and easy to use.

  3. Stack

    What I’ve worked with.

    Working with agents
    Claude Code, Cursor, Claude API, OpenClaw, Hermes Agent
    Backend
    Python, FastAPI, Pydantic, pytest, TypeScript, Node.js, Fastify
    Databases
    PostgreSQL, pgvector, SQLAlchemy, Drizzle, Alembic
    RAG and semantic search
    RAG, sentence-transformers, pgvector, PyMuPDF
    APIs and integrations
    REST API, Webhooks, WebSocket, Telegram Bot API
    Frontend
    Next.js, React, Tailwind, shadcn/ui, SvelteKit, three.js
    Blockchain
    Solana, Rust, USDC
    Infrastructure
    Docker, Docker Compose, Caddy, VPS
    Automation
    n8n
    Earlier
    Kotlin, Java
  4. What I can build for you

    All of it already runs in my own products.

  5. AI agents and automation

    Agents that sign up on their own, take tasks and follow a protocol. Payouts in cycles, smart-contract checks before a trade, scheduled messages. I’ve also worked with n8n.

  6. RAG and semantic search

    Scientific papers are cut into passages, a local model turns each into a vector, and pgvector in PostgreSQL finds what matches by meaning. That is how OpenSolve agents get the right passages from papers.

  7. Telegram bots

    Conversation, memory of the person, onboarding, and messages the bot sends first. Telegram sign-in, a Mini App.

  8. Backend and integrations

    FastAPI and PostgreSQL, auth, database migrations. Scientific databases (PubMed, arXiv, Crossref), market data (DexScreener, GeckoTerminal), sign-in with Clerk, crypto wallets.

  9. Dashboards and landing pages

    Screens like the OpenSolve Observatory: live figures, a graph of tasks, a feed of events. After four rounds of work, the ShadowJung landing page beat the well-known design sites Astral Frontier and Persepolis Reimagined in a blind comparison.

  10. Tests

    926 tests in OpenSolve, 32 for the smart contract, plus an audit report.

  11. The path

    At Sber and Yandex I answered small-business owners whose tools had stopped working. Good training: you see a product through the eyes of someone who is annoyed with it. Then key accounts at PIK, a property developer, and a year in a foundry.

    • Now · OpenSolve, ShadowJung, XENEURO.
    • Freelance and my own products · Software engineer: sniper bots that buy new coins at launch, trading bots, mini-games for crypto projects, software for blockchain operations
    • Metallurg-Tulamash · Fettler in the foundry
    • PIK Group · Key account manager
    • Yandex · Technical support specialist
    • Sber · Technical support specialist
    • KARI · Customer support, written enquiries
    • MegaFon · Product promotion specialist
    • Tula State University, incomplete degree

Message me on Telegram

@riviaConfiglomov_rh@outlook.com

Russia and Asia · remote only · GitHub · X

Artur Solodovnikov · AI developer