ALEKSANDR SHULGIN · AI & BACKEND ENGINEER · GMT+10 · AVAILABLE FOR ROLES AND PROJECTS

Codefor anytask.

Describe what you need in plain words and get back a system that runs without me: bots, automation, AI agents, trading systems.

What brings you here

★ §02 / Work

What I've built.

Flagship · trading infrastructure

Market-Making on Prediction Markets

Trading infrastructure on two exchanges, running unattended in normal operation: liquidity quoting, treasury, positions and risk, alerts and a manual stop. Custom low-latency infrastructure and a research loop, the whole stack on me.

×13
lower order latency
sub-50ms
p99 order submit
24/7
Unattended
3.9M
Trades analyzed

Order submit 290-596 → 18-41 ms, March 2026. Numbers: my own measurements on my own infrastructure, 2025-2026. Methods are closed on purpose.

Subsystems05
S1

Execution layer

A series of trading bots: async Python orchestration with Rust on the hot path. Sub-50ms order submission, a 13x latency gain.

PythonRustasync
S2

Market-making engine

Liquidity quoting for a rewards program, careful position reconciliation, its own treasury. Around the clock, with alerts and a manual stop.

CLOBtradingtreasury
S3

Edge research

Market research on our own data: 3.9M trades across 6,249 wallets.

analyticsEVMdata
More work
Products, tools, open source
// content platformContent Factory2026PRODUCTION

A content-generation platform for several channels. A FastAPI backend and a Vite frontend, multi-provider LLM orchestration with fallback, a worker queue, scheduled publishing. Running 24/7.

PythonFastAPIViteLLM
Type web+LLMStatus live
// tool for businessSite Check2026PRODUCTION

A free express check of a website against Russia's personal-data law (152-FZ): foreign services, personal data collected without consent, a missing policy. Runs in an isolated Cloudflare Worker: the visitor supplies the address, our servers stay out of it. First rung of the ladder: a free check, then a paid audit, then fixes.

Cloudflare WorkersJavaScript152-ФЗsecurity
Type toolStatus live
open
// graphics engineLiving Canvas2026PRODUCTION

A soft-3D engine for living characters on plain canvas 2D: a parametric mesh, Lambert lighting, perspective and emotions - zero libraries, 60 fps. Live demo: the ghost right in this section, go click it.

TypeScriptCanvas3Danimation
Dependencies 0Status live x2
open
// devops toolSetup Manager2025PRODUCTION

A web tool for managing environment config and deployment: environment presets, settings sync, fast bootstrap on a new machine. Public site on a custom domain.

Node.jsJavaScriptCSSdeploy
Type publicHost Cloudflare
// product · macOSQwerty Switcher2026PRODUCTION

A native macOS auto-switcher for keyboard layouts: key interception via CGEventTap, language detection with a dictionary (≈697K words: a Bloom filter with exact confirmation) plus n-grams, replacing the typed word on the fly. Learns from the user’s corrections and detects games so it never touches keys mid-game. Free, no subscription; stays offline unless you opt in to signed update checks. Source is open on GitHub. 1310 automated checks.

SwiftAppKitCGEventTapSwiftUIproduct
Platform macOS 13+Model free · no subscriptionSource github.com/Sanexxxx777/QwertySwitcher
open
Google DeepMind · formal-conjectures

My code is in Google DeepMind's repository

Five formalizations of Erdos and Green problems written in Lean 4, reviewed by Google DeepMind maintainers and merged into their open formal-conjectures repository; one more merged PR links an external proof, six in total. Lean checks the formal proofs: 0 sorry, clean axioms. Next to it, an evolutionary search that reproduced a numerical bound at AlphaEvolve level. ∎

5
Formalizations merged into DeepMind
M1-M5
min-overlap values proved
99.5%
match to the SOTA bound
0 sorry
Lean kernel, clean axioms

Every card links to the PR or the repository: verifiable in a minute.

Contributions08
Client work
// client workTransaction Guardian2026CLIENT

A client service for monitoring on-chain transactions. Event subscriptions, filters by address and signature, Telegram alerts, a stats dashboard. Delivered to the client and launched in their production.

PythonWeb3EVMTelegram
Type clientStatus delivered
// client workHub Research2026CLIENT

A client AI assistant for crypto research. It gathers data from many sources, distills it into structured reports with LLM analysis, and delivers them through a bot. Delivered to the client and launched in their production.

PythonLLMGeminiresearch
Type clientStatus delivered
// client workTwitter Intel2026CLIENT

X/Twitter monitoring across crypto projects with AI summaries, importance scoring and noise filtering. Telegram alerts with inline controls. Delivered to the client and launched in their production.

PythonLLMTelegrammonitoring
Type clientStatus delivered
Retired
// deep analyticsCrime Token Scan2026RETIRED

A Telegram bot for deep crypto-asset analytics: a custom pipeline that collects and scores signals, an AI comment on every event, clean visuals. Ran in production 24/7 from May to September 2026, now retired; kept here as an example of a full build.

PythonLLMTelegramanalytics
Type publicStatus retired · Sep 2026
// our own service25.VPN2026RETIRED

A VPN service on custom infrastructure: the site, an account area, key issuing and renewal, a Telegram bot, multi-server rollout, protection against sharing and leaks. It ran with paying users; it is no longer sold and stays here as an example of a full build.

PythonLinuxnetworkTelegram
Type serviceStatus retired
// web appSector Map2025RETIRED

A crypto-market web dashboard with interactive heatmaps by sector. A Python backend aggregates CEX/DEX data; the frontend is a lightweight SPA in vanilla JS + D3. Ran on a custom domain via Cloudflare Tunnel until June 2026, now retired.

PythonJavaScriptD3Cloudflare
Host self-hostedStatus retired · Jun 2026
// monitoringNFA Dashboard2026RETIRED

An infrastructure-monitoring panel: server and service status, metrics, alerts, a daily digest in Telegram. 7 tabs, JSON persistence, cron pushes. Retired in June 2026; the demo repository stays open.

Node.jsJavaScriptPM2cron
Tabs 7Status retired · Jun 2026
open
Stack, through the work

click a technology to open the work where it is used

★ §03 / Sites

Sites I've shipped.

What clients say

Translated excerpts from clients' Telegram messages after handover.

Serzhantovo Horse Yard: preview
"You really did a huge job for me. People say from the outside that the site is really cool. I do not notice such details myself, but they do. Thank you so much."
Alyona, owner of the horse yard
open the site ↗
Mechta Tut anticafe: preview
"Now this is the perfect site, really. A hundred out of ten. Everything is clean, everything is beautiful. I will present it everywhere: Telegram, VK, YouTube, pinned in every video. Thank you for taking it on."
Vasily, Mechta Tut anticafe
open the site ↗
01Serzhantovo Horse Yardclient site2026
Serzhantovo Horse Yard: preview

A horse yard and petting farm: animal pages, services, events, gallery. The owner edits content directly - a one-time email code instead of passwords or a hosting panel; a save publishes JSON that the page picks up without a rebuild. Uploaded photos are resized and converted to WebP automatically, and every save writes a backup.

Next.jsPHPCMSSEOhorsesfarm.ru ↗
02Tatyana Bunina - Hoof Trimmingclient site2026
Tatyana Bunina - Hoof Trimming: preview

A site for an author's course on physiological hoof trimming: a fifteen-lesson programme, tariffs, videos, sign-up. Built from scratch rather than exported from a site builder - no third-party scripts or trackers. Content is edited through the same editor as the horse yard.

03Mechta Tut anticafeclient site2026
Mechta Tut anticafe: preview

An anticafe and board-game club in Saint Petersburg: event rooms with a rental calculator, pay-per-minute visits, a 3D carousel of the game collection, a Telegram request flow with no server-side forms. The owner edits prices and announcements in an admin panel; the page picks them up without a rebuild.

HTMLCSSJSadminme4tut.ru ↗
04Make It Beautiful: the studio's site-filmour own site2026
Make It Beautiful: the studio's site-film: preview

A site that moves like a reel: a pen draws a ring, the visitor closes it, and their own circle cracks the page into a film strip, a phone, slides and a logo. Eight worlds, two gestures, sound after the first touch, Russian and English. Drawn by code on one canvas, no images and no heavy libraries.

TypeScriptCanvasGSAPWebAudioshulgin.is-a.dev/studio ↗
05Burn the Letterour own site2026
Burn the Letter: preview

A ritual for anger: you write a letter on a ruled notebook sheet, strike a match on the box and touch the paper. The fire runs up from your finger, the words flare before the paper, char and holes stay behind the rim. Then three breaths with the last coal and a card with an ash print. The text never leaves the page: an automated test captures every request and looks for the letter's words. The fire is a hand-written WebGL2 shader with no libraries; the desk holds a cassette player and a tablet with radio.

TypeScriptWebGL2WebAudioWorkersshulgin.is-a.dev/burn ↗
★ §04 / Lab

Live mechanics.

Three of thirty-four. Each one runs right here, each has its own address and source.

Torchere with a pull cordPull the cord
ru
Dot-matrix glyphsMove over the dots
ru
Cursor spotlightMove the light
ru
The whole lab
★ §05 / Wow site

Flood any website.

Two phones: a site before the flood and half under water

CLOUDFLARE WORKER · BROWSER RUN · MATTER.JS · CANVAS 2D

Type any website's address, hold your finger, and the page goes under water: headings come loose and float, images sink, tilting the phone rocks it all. Share the picture with anyone.

  • A real browser renders the site inside Cloudflare's isolated network; none of my servers are involved
  • Matter.js physics: three-point buoyancy, spring-line waves, 60 fps on a phone
  • One evening from idea to live link: Cloudflare Worker, a KV cache of levels, 32 automated tests
Flood a site
★ §06 / Films

Films rendered from code.

Evidence over Claims3:57

Evidence over Claims

eval, micro1 hackathonen

An eval built for the micro1 Eval Hackathon: a pipeline run says COMPLETE, the logs show which tasks failed. Seven 2026 models got every item right; Llama-3.1-8B and Gemma-3-4B scored 2 and 3 %. AI voice-over.

one gold evidence dot carries every scene

A website, end to end13s

A website, end to end

serviceru

Three pieces of work, how an order runs, and what the client keeps. Each piece is labelled on screen: the first two are our own reference builds, the third one runs for a real client.

bright springy montage

Serzhantovo Horse Yard17s

Serzhantovo Horse Yard

client workru

A film about the place and the site built for it. The client's own photographs; opening hours and wording checked against the live page, including the rules for approaching the horses.

editorial spread

★ §07 / Method

How I work.

Scope

Agree on done first

Before I write code we agree on what done means: inputs, outputs, limits and how we will check it.

24/7

I run it in prod, not in a demo

A task is done when it's in production. Not "runs locally" but actually holds traffic, survives restarts, recovers from network drops.

Numbers

I measure, not guess

"Feels slow" is not an argument. Benchmark, profiler, numbers. No measurable gain means the optimization isn't needed.

Handover

Handover, not hostage

You get the code, the access and a short runbook: what runs where, how to restart it, what each alert means.

★ §08 / Timeline

What I've shipped.

v2026.09September 2026

Fifth formalization accepted into DeepMind

Research

Erdos 885, the k = 4 case: proved in Lean 4 with an explicit witness from a computer search, a formalization of Bremner's 2019 result. PR #6509 was merged into google-deepmind/formal-conjectures on September 23, 2026; PR #4379, linking an external formal disproof of problem 90, was merged the same day. Six PRs merged in total, one more open.

v2026.09September 2026

Qwerty Switcher 0.11.0: free, source open

Production

The native macOS layout switcher went free with no subscription, the source is open, updates are signed and opt-in. 872 automated core checks.

★ §09 / Contact

Let's work together.

Open to AI/LLM and Python backend roles. About two years of hands-on work: my own products in production, trading systems and client projects. I pick up the unfamiliar fast and ship it.

Reply within a day.

Write on Telegram
Business orders are always open: websites, bots, integrations. Plain language, a contract and a receipt.Describe a task
▌ Format
Full-time · contract · freelance

Remote or on-site. Open to relocation: EU, Asia, UAE.

▌ Focus
AI / Backend / Automation

Python backend, LLM integrations and agents, bots and automation, crypto/web3 data, websites.

▌ Languages
RU · EN

Native Russian. English: docs, code and written communication with ease; happy to do a screening call.