Independent systems builder · Minsk, works remote Taking work · paid in USDC / USDT

Mikhail
Yakimushkin

I build small production systems end to end — and I run them. Not prototypes, not demos: services that stay up, get used daily, and have someone on call. That someone is me.

Trading engines, Telegram bots, finance apps, internal CRMs, self-hosted infrastructure. I work solo and use AI as the workforce, which is why one person ships at the pace below. Every number on this page comes from a repository you can ask me to walk you through.

9 systems built
and deployed
110k lines of
shipped code
877 commits in
this workspace
1 person on
every one

Selected systems

Status as of Aug 2026
Stack Next.js · SQLite
TypeScript · PWA
Scale 273 files
49,100 lines
Status In daily use

Tally Web

A personal finance manager built to replace Money Pro — double-entry ledger, budgets, recurring transactions, reports, installable as a phone app with push notifications.

The interesting part is the ledger. Money has to balance to the cent across accounts, transfers, splits and currency conversion, and it has to keep balancing after every edit and every migration. 15 Playwright end-to-end specs guard the paths where money moves. Self-hosted on my own hardware behind Tailscale, with systemd timers driving the morning and evening notifications.

Stack Python · React
Bybit API
Scale 12,200 lines
107 test files
Status Running live

AlgoDesk

A trading terminal for crypto perpetuals that takes a strategy through the same engine three times — backtest, paper, live — so the numbers you approve are the numbers you get.

One engine, three modes, no forked code paths: that constraint is what makes a backtest trustworthy, and it is where most retail trading systems quietly cheat. Position sizing, per-trade and portfolio risk caps, cooldowns and a drawdown kill-switch are enforced in the same layer whether the fill is simulated or real. 107 test files exist because a rounding error here costs money.

Stack Python · OKX API
Telegram Bot API
Scale 63 files
8,800 lines
Status Running live

Grid Supervisor

A supervisor for grid trading bots on OKX. When a bot finishes its cycle, the system re-runs the backtest on fresh data, sends a decision card to Telegram, and starts the next bot when you tap the button.

It is a closed loop with a human in exactly one place — the approval. Everything either side of that tap is automated, including the part most people skip: re-validating the parameters against what the market has done since the bot was configured.

Stack SwiftUI · Swift
macOS native
Scale 232 files
30,600 lines
Status Shipped

Tally for Mac

The native desktop app the web version grew out of — a full double-entry ledger in SwiftUI, built to feel like a Mac app rather than a website in a window.

Proof that I do not need a browser to build an interface. Keyboard-first data entry, a real document model, and an accounting core with its own test suite underneath the UI.

Stack Python · LLM agents
scheduled workers
Scale 34 files
7,000 lines
Status Weekly cadence

Council

An agent system that reads across my projects and my working week looking for recurring friction, then issues five decision cards every week — each one a specific change with a reason attached.

The engineering problem is not calling a model, it is trusting the output: deduplicating findings across runs, refusing to merge unrelated items into one card, and telling silence in a data source apart from a broken sensor. Those three failures are what make most agent systems get quietly ignored after a month.

Stack Python · Ollama
local GPU · systemd
Scale 27 files
3,000 lines
Status In daily use

Charge

A Telegram coach that runs entirely on my own hardware — a local model on an RTX 5070, no data leaving the house.

Built around one rule: the bot never fails because the model is down. If inference is unavailable it degrades to templates and buttons and the user never sees an error. That graceful-degradation pattern is what I bring to any product with an LLM in the critical path.

Stack FastAPI · Whisper
Tailscale · Swift
Scale 0.2–0.5 s
per transcript
Status In daily use

Dictation

Press a hotkey on the Mac, speak, press it again — the text appears at the cursor in whatever app is in front. Recognition runs on the GPU in my home office.

Audio travels over a private SSH tunnel on a Tailscale network, nothing is exposed to the internet and nothing is stored. Sub-half-second round trip, which is the difference between a tool you use and a toy you abandon.

Stack Next.js · CRM
document pipeline
Scale Client-facing
business system
Status In production

CertFlow

A CRM and document pipeline for a product-certification business: deals, clients, regulatory documents, inspection control and risk tracking, plus the public website and an automated lead worker.

The one on this page with real users who are not me, and real deadlines attached to regulatory filings. It generates compliance documents from structured data — the kind of work where a wrong field is a rejected application, not a cosmetic bug.

Stack Stalwart · mbsync
DNS · Linux
Scale Full migration
own domain
Status Delivered

Mail Server

Migration off a hosted provider onto self-run mail on our own domain — full archive extraction, a mail server in the office, and a webmail front end on a droplet.

Included because email migration is where infrastructure claims get tested: you get one attempt with someone's entire correspondence, and deliverability either works or it does not.

How I work

Four commitments
Deployed, not demoed

A handoff includes the thing running on your infrastructure, with process supervision, logs and a restart story. A repository is not a delivery.

Tests where money is

I do not test everything. I test the paths where a mistake costs something — balances, orders, payouts, document fields — and I tell you plainly which paths those are.

AI as leverage, reviewed by me

I use AI aggressively and I read every line before it ships. You get the speed of it without inheriting its mistakes — that division of labour is the whole reason one person can deliver this list.

Fixed scope, fixed price

I quote a defined deliverable with a date, not an hourly rate that drifts. If the scope changes we re-quote, in writing, before I build.

What I can build for you

Typical engagements

Telegram bots & automations

Intake bots, AI assistants, document parsing, scheduled reports, CRM integrations. Deployed on your server with monitoring, not left in a notebook.

1–2 weeks · fixed price

Trading & market tooling

Exchange integrations, backtest harnesses, risk and position sizing engines, supervisor loops, dashboards over live positions.

2–4 weeks · fixed price

Internal web apps

Next.js and SQLite or Postgres, authentication, roles, reporting, installable on phones. The tool your team actually needs instead of the spreadsheet.

2–4 weeks · fixed price

Agent pipelines for dev teams

Setting up AI coding agents that hold up in a real repository: tool servers, review pipelines, scheduled agents, and the guardrails that stop them producing confident nonsense.

Workshop + implementation

QA on live products

Structured bug hunts against a running app — API contract checks, data-integrity scans, performance measurement, reproducible reports your engineers can act on.

Per engagement

Self-hosted infrastructure

Getting off rented SaaS: Linux services, private networking, GPU inference, backups and the boring reliability work that keeps it running.

Scoped per project

Start with a scoped conversation

Tell me the outcome you need and the date you need it by. I will come back with a fixed scope, a fixed price and a delivery date — or tell you honestly that I am the wrong person for it.

Telegram — @leshwas yakimushkinm@gmail.com github.com/yakimushkinm-prog

Dictation is public — the rest stay private, because client systems and my own finances live in them. I walk through any of them on a screen share, and I will give a read-only invite once we are working together.