ArtJeck is the studio practice of one independent AI engineer. I design, build, and test custom web, mobile, and AI products — done right, and you own the code outright.
I matched a carrier's official tax filing to the penny.
An AI filing pipeline I designed and built, backed by 562 automated tests and running in production for a real client.
562
automated tests
Exact
filing match
Live
in production
From AI agents to app-store releases — and on every one, you work directly with the engineer building it.
AI that holds up in production, not just the demo.
Custom agents and assistants that answer from your own documents and take real work off your team's plate. Built with evaluations and guardrails, so accuracy holds when real customers are the ones asking.
Fast, custom sites and web apps you actually own.
Websites and web apps built clean, quick, and exactly to your brand — no templates, no monthly fees holding your site hostage. The code is yours from day one, built to grow with you.
Mobile apps from first idea to the App Store.
Native iOS apps in SwiftUI and cross-platform builds that feel at home on every device. I take you from a focused MVP to a tested production release — your app, your code, your roadmap.
Catch the bugs before your customers do.
Independent functional, regression, and edge-case testing on software you or another team built. You get a clear written report of what breaks and why — a senior second set of eyes before you put your name on it.
Your business data in one place, finally connected.
Databases set up, migrated, and connected — spreadsheets turned into real systems, your tools talking to each other through their APIs, and messy documents like PDFs and invoices extracted into clean, structured data.
Get found on Google — and know what's working.
Technical SEO that makes Google actually understand your site: structured data, local pages, speed, and clean indexing. Plus analytics done right — GA4, Search Console, and a dashboard that shows where customers really come from.
Each one is live, tested, and open to inspect — the actual deliverables, not mockups. I lead with the outcome, then show how it was built.
Trucking carriers spend hours every quarter reconciling fuel and mileage data, hand-typing per-state lines into the gov portal, and second-guessing whether the math matches what the state will recompute.
An end-to-end pipeline that ingests raw mileage and fuel files, computes a state-portal-ready return with exact CDTFA math, and runs an AI agent over it to flag missing surcharges, MPG anomalies, and audit-bait patterns before filing.
Python, pandas, openpyxl, pdfplumber, Anthropic Claude, multi-tenant client registry, per-truck Excel deliverables
562 automated tests, including a real-data backtest that matches a Kentucky carrier's official CDTFA filing to the penny.
A real client (DM Express, KY) files quarterly through the pipeline, with an AI review note and concrete next steps before each filing.
A resale business sells the same inventory across many marketplaces at once. Listing each item by hand on every platform — and keeping prices and 'is it still listed?' in sync — is slow, repetitive, and error-prone.
A cross-posting engine that describes an item once in a normalized model and publishes it to every marketplace through one uniform adapter interface — with bounded retries, fast-fail on rejections, and idempotency so a scheduled re-run never double-posts. Anything irreversible, like repricing a live listing, stops for a human approval first.
Python, adapter-pattern marketplace clients, deterministic listing generation, SQLite listing ledger, retry/backoff orchestration, contextvar-scoped execution grants for human-gated actions; the production system spans eight marketplaces behind private adapters
The public clean-room showcase ships 30 tests over the retry state machine, permanent-vs-transient handling, idempotent skips, partial-failure reporting, the deterministic listing rules, and the approval gate.
Running for a live, revenue-generating client (anonymized). A sanitized showcase of the architecture is public and runs end-to-end on synthetic data; the production code and integrations stay private.
Most AI 'QA agents' re-audit the whole repo every run, re-report the same findings until you stop reading, call flaky and stale tests alike 'failures', and sign off with an LGTM. And once an agent is writing the code, a loop where it also tests its own work has no independent gate left in it.
A QA agent with a stored baseline, so a repeat run is a delta — NEW, REGRESSED, STILL_OPEN, RESOLVED — with regressions ranked first and every open finding carrying its age. It has no Edit tool by design, quarantines flaky tests only with an expiry date attached, requires a citation before it will write a failure off as a stale expectation, and closes on one of four verdicts where 'blocked' is a legitimate answer. Every line a finding cites is hashed, so the next run knows which evidence the code moved and re-verifies exactly that.
Zero-dependency Claude Code plugin, 10 slash commands behind one front door (/verdict:run picks baseline, delta, or a scoped review itself), the same doctrine as five skills for Cursor, Codex, and any other coding agent, a Python MCP server exposing 10 read-only tools, a fact harness that measures counts and diffs so the model contributes judgment only, an exit-code release gate for CI, a local tier that runs unattended nights on an 8B model with zero Claude tokens, a 24-technique test-design catalog, 5 report templates, and six hooks — write-scope, Bash-scope, and a Stop hook that blocks a run which tried to hand-write its own state. MIT.
Ships with its own eval: seven scored fixtures — seeded defects, a delta run against an authored history, a TypeScript twin, root-cause with a decoy, a spec review with no code, AI-generated slop, and an adversarial repo whose suite prints ALL TESTS PASSED while exiting 1 — plus a key it did not write: 40 SWE-bench Verified issues, the maintainers' fix located in 36 of 37 scored runs. Published results include the misses and the control: plain Claude Code located those defects just as well in a twelfth of the time, and the repo says so next to the 8/8s. 1,338 tests and a scorer-regression corpus run in plain CI with no model.
Public and MIT, 73 tagged releases in its first month. Runs unattended nightly against a production codebase — its first fully unattended run refused to execute the suite because a live .env sat in the checkout, said so, and still delivered a delta report, then found a leaked API token untracked in the repo. Pointed at strangers' repositories, it found a rotate_file in boltons that, at keep=1, deletes the file it rotates and a FilePerms bug that had lived 4,595 days, both reported upstream. Turned on itself, it filed the defect that made its own anti-fabrication check imitation-proof: every run now signs the run history with a hash chain a copied state cannot reproduce.
A UK Community Interest Company researching ancient Eurasian steppe heritage needed a public face serious enough for museums, universities, and grant panels — and needed to maintain it themselves, without a developer on retainer. A CMS answered the brief; a five-service stack on a small VM, patched and paid for by volunteers to serve content that changes a few times a month, did not.
Delivered in two phases. First a bespoke WordPress build — content types in a custom plugin, design in a custom theme, the whole environment provisioned by one idempotent script. Then a static rebuild that kept everything visible and discarded everything else: content exported from the REST API to JSON, the theme's CSS carried across, original URLs preserved down to the trailing slash, and retired pages redirected rather than broken.
WordPress with custom gss-core plugin and gss-journal theme, Docker Compose, Caddy; rebuilt on Next.js 15.5 static generation with three runtime dependencies, JSON content, self-hosted Latin and Cyrillic typography, HMAC-signed contact tokens, Resend, Vercel
Verified the paths that had been failing silently: the contact form now renders only when delivery is configured and surfaces real errors instead of a false success, the spam gate rejects without telling a spammer why, and the hero is correct with JavaScript disabled and honours reduced-motion preferences.
Live at greek-scythian.org — 23 fully prerendered pages, no database and no server to maintain. The WordPress phase is public under GPL-2.0. Every original URL still resolves, and the largest image on the page ships at 120 kB instead of 361 kB.
Every AI portfolio claims agents; almost none lets a visitor talk to one. A public demo of a personal RAG system is also a security exercise — the corpus, the port, and the inference budget are all attack surface, which is why most demos are either locked behind a form or open and leaky.
A public slice of my personal RAG assistant, live on this site at /lab/brain. It answers questions about my services with citations to the documents it retrieved, and declines what its curated corpus does not cover — including its own infrastructure, by design. Inference runs on local models in my home lab at $0 per request, behind a Vercel zone, a service-token proxy, and a Cloudflare Tunnel.
Python, FastAPI, Qdrant / Chroma hybrid retrieval, an OpenAI-compatible gateway to a local model, an MCP server (stdio + token-gated HTTP), Next.js 16 multi-zone on Vercel, Cloudflare Tunnel, Turnstile, per-visitor rate limiting at the origin
141 automated tests plus an eval harness over a synthetic regression corpus — retrieval hit-rate, grounded-answer checks and abstention cases — and a live regression test that asks the deployed agent about its own infrastructure and requires the refusal.
Open source and in daily use, exposed over MCP so Claude Desktop and Claude Code query and teach it directly — and public at /lab/brain, where a visitor gets cited answers from a local model, a labelled recorded-answers fallback when the lab sleeps, and a kill switch that is one environment variable.
A jewelry brand needed a real online store — catalog, reviews, promotions, and self-service management — without a monthly platform fee or a dashboard to learn.
A full Next.js storefront with a product catalog, image galleries, reviews, stock, a Buy-Now flow, and token-based personal offer links — plus a secured admin where the owner manages everything themselves.
Next.js, React, TypeScript, Prisma + libSQL (Turso), Vercel Blob uploads, Tailwind CSS, TOTP two-factor admin
Admin gated behind two-factor auth, validated image uploads, and cached queries for fast product pages; storefront and admin flows tested end to end.
A live store the owner runs themselves — add products, publish reviews, send personal offers — with no recurring platform cost.
A small trucking company needed a credible, fast, phone-friendly site to help recruit drivers.
A one-page React/Vite site with light and dark themes, animated sections, and a structured driver application flow.
React, TypeScript, Vite, Vitest, mailto application flow, Vercel
29 tests using equivalence partitioning, boundary value analysis, decision tables, and state transition testing.
A real client site tuned for iPhone, accessibility, and a sub-500KB first load — now their front door for recruiting drivers.
Operating a multi-machine AI lab from anywhere usually means SSH and a raw shell — powerful, but reckless to hand to an autonomous agent.
An MCP server that gives any agent safe “hands” on the lab: health checks, model management, free local inference, and a deliberately gated remote shell — allowlist only, no shell metacharacters, hard timeouts.
Python, MCP (stdio + token-gated HTTP), httpx, subprocess argv (no shell=True), launchd, Tailscale
A pytest suite covering the command-gating safety logic, runnable offline.
Open source and running around the clock; an agent can operate the lab from an iPad while the dangerous operations stay locked behind validated tools.
Daily inbox review was noisy and manual, with important messages mixed into low-priority mail.
A local agent that pulls mail over IMAP, uses a local model to triage and summarize, then sends a Telegram digest.
TypeScript, Node, imapflow, mailparser, Ollama, SQLite, Telegram Bot API, launchd
Test-connection and dry-run paths, SQLite dedupe for idempotency, and cost reporting per run.
A self-hosted morning digest that runs on schedule, entirely on local models at $0 API cost.
Bills of lading are still often reviewed and keyed by hand, which slows operations and creates avoidable data-entry risk.
An in-progress extractor that turns BOL documents into structured shipment data with validation and review-ready output.
OCR / document parsing, LLM structured extraction, JSON schema validation, field checks, export workflow
Planned around sample BOL fixtures, required-field checks, edge-case documents, and regression tests for extraction quality.
In progress: designed to cut manual BOL entry and make shipment data easier to review, reuse, and automate.
Premium pricing buys senior hands, real testing, and code you own outright — so you pay once and keep what you get.
The person who scopes your project is the one who builds it. No account managers in the middle, no junior developers learning on your budget.
Scope and price are agreed before work starts. You know what you're paying and what you're getting — no surprise invoices halfway through.
Testing isn't an afterthought bolted on at the end. It's how I work, so bugs surface in my hands and not in front of your customers.
All the code, all the accounts, all yours. No lock-in, no licensing traps, no monthly fee to keep your own product running.
Quoted verbatim, with permission, from the businesses whose operations run on this work.
Eugene has been taking care of the technology at my companies — Hills Freight, Force Trans, and Impala Freight — for years. Computers, software, the network, whatever stops working. When something breaks he picks up the phone and gets it handled, and in trucking, downtime costs real money, so that matters to us. He set things up so the office just runs, and he explains everything in plain language instead of tech talk. Honest, reliable, easy to work with. I'd recommend him to any business that needs somebody to take care of their IT.
The biggest thing he built for us is the system that files our quarterly IFTA fuel tax. It used to take me about an hour of stressful work right against the deadline, and it was easy to make a mistake. Now it takes about three minutes. He built in an AI review step that checks every return before it goes out, and it caught errors we had been making for years without knowing it. He's reliable, he explains things in plain English, and he does what he says he'll do.
Six steps you can watch happen — so you always know what's being built, what it costs, and what comes next.
I learn your business, your users, and what success looks like before writing a line of code.
I design the system and we agree on scope and a fixed price, so there are no surprises later.
I build in clean, documented code that's yours to keep, and keep you in the loop the whole way.
I test against the edge cases and write up exactly what I found and fixed — before your customers see it.
I ship it live, hand over every key and credential, and make sure it holds up under real use.
I stay close after launch to refine, extend, and grow what we built together.
I treat design, build, and testing as one job, not three handoffs — because the seams between them are where most software quietly breaks. I'm the same hands from sketch to ship, so the details line up.
Testing discipline is my edge. On one AI project — an IFTA tax-filing automation — I wrote 562 automated tests and ran a real-data backtest that matched a Kentucky carrier's official filing to the penny. I'd rather find the problem in a test than have you find it in production.
Working with me is direct — and “me” is Eugene Menshikov, an independent AI engineer in Sacramento with a software-quality background. You talk to the person doing the work, you get honest answers about what's hard and what's worth it, and you leave owning software you understand.
Senior hands on your project, start to finish
Built to be maintained, not just launched
Tested against real-world edge cases
Plain answers and a clean handoff
Your AI can run on a private, self-hosted stack I build and operate — free local models by default, paid cloud only when it earns its place. For you that means lower running costs, real data privacy, and no surprise bills.
Every AI call runs through one gateway with automatic failover and a hard spend cap — so an automation can't quietly run up a bill against you.
Always-on machines run free local AI for chat, code, and search, with a vector index and database holding your project's knowledge — no per-token meter ticking.
Everything sits on a private network — nothing public — and the services stay up around the clock, so what you depend on is there when you need it.
Private by default — nothing is exposed to the public internet.
Design, engineering, and quality assurance from one engineer — the seams where projects usually fray simply aren't there.
AI that answers from your own documents and runs your workflows — on private infrastructure with local models, automatic failover, and hard spend caps, so it stays accurate, private, and predictable.
Fast, custom websites, web apps, and mobile apps in modern, maintainable code that's entirely yours to keep.
I test the way your customers will, then harder — functional, regression, and edge-case coverage with clear written reports.
On-site around Sacramento, in the Bay Area when it helps, and remote for everyone else.
Agents, RAG assistants, and workflow automation for local operators, logistics, and compliance work.
See the workPrototype-to-production AI: agents and assistants hardened with evals, guardrails, and launch QA.
See the workOperations automation — intake, routing, reconciliation, reporting — for teams across the region.
See the workThe local repair practice: on-site across Sacramento and seven nearby cities, or remote.
See the workThe full picture is on the AI automation services overview.
The things people want to know before a first call, including the ones with awkward answers.
Eugene Menshikov, an independent AI engineer based in Sacramento, CA, with a software-quality background. The person who scopes your project is the person who designs, builds, and tests it — there are no account managers, no juniors, and no handoffs between teams.
AI automation projects start at $8,000 as a fixed quote, and advisory work is $150 an hour. Websites start at $1,500, mobile apps at $7,500, and independent testing at $3,000. Every price on this site is published rather than hidden behind a contact form, and you get the number in writing before any work starts.
An agency sells capacity: several people working in parallel, and cover when someone is away. An independent engineer sells accountability and depth: one person holds the whole system, so nothing is lost in translation. If you need multiple workstreams at once, or your procurement requires a vendor with more than one person, hire the agency — that is a real requirement and I will tell you so on the first call.
It is the point, not an add-on. Every system ships with automated tests and, where a language model is involved, evaluations that run like tests. The IFTA filing agent carries 562 automated tests and a backtest that matches a real Kentucky carrier's official state filing to the penny. AI fails quietly rather than loudly, so this is the difference between a system you can rely on and a demo.
Yes. Sacramento is the base and where on-site work is easiest, Bay Area projects get on-site time when it helps, and everything else runs remotely. Clients outside California are normal — the trucking carrier running the IFTA agent is in Kentucky.
Entirely. The repository, the accounts, and the documentation are handed over, with no monthly platform fee and no dependency on me to keep it running. You can hire someone else to maintain it, which is exactly the point of writing it to be maintained.
The workflow that repeats on a schedule, follows written-down rules, takes structured files rather than conversations as input, and where mistakes are both expensive and detectable. A process that runs twice a year, or whose rules live in one experienced person's judgement, is a bad first project no matter how tedious it feels.
A free intro call of about 30 to 45 minutes, where you describe the problem and get an honest read on whether it is worth building and roughly what it costs. If it is a fit, a fixed written quote follows before any work begins. If an agency or a different specialist suits you better, you will hear that instead.
Tell me what you're building and where it's stuck. I'll tell you straight whether I'm the right engineer for it — and what it would take to do it right.
Work with me if you
