I work spec-first — requirements → design → build → test → deploy — so what I ship survives contact with real users. That QA discipline is the part most AI work skips, and it's my edge. Everything below is running, tested, and readable.
🌐 artjeck.com · ✉️ hello@artjeck.com · 💼 LinkedIn
- Production LLM pipeline, live: ifta-agent files real quarterly IFTA fuel-tax returns for an interstate carrier — 555 automated tests, backtested against a real state filing to the penny, ~$0.10 model cost per reviewed return. Its LLM review agent (17 grounded tools) caught filing errors the manual process had been making for years.
- Published AI QA agent: verdict — baseline → delta runs, a fact harness (the model judges, the system measures), signed run history, an exit-code CI gate. 43 releases, 419 tests, six scored eval fixtures with the misses published alongside the 8/8s.
- Human-in-the-loop by architecture: an 8-marketplace commerce engine where irreversible actions are structurally gated behind human approval — no approved grant, no code path to the write.
- Stack: Python · TypeScript · FastAPI · Next.js · Anthropic SDK · MCP (3 servers built from scratch — one listed in the official MCP Server Registry) · RAG/Qdrant · Ollama/LiteLLM self-hosted lab.
- Open to: AI engineering · AI automation · QA — remote (US), Sacramento, or SF Bay Area. Also available for freelance builds via ArtJack Technology.
| Project | What it is |
|---|---|
| ifta-agent | Production IFTA fuel-tax filing pipeline for interstate carriers. Deterministic math computes every number on the form; an LLM review agent over 17 grounded tools audits each return against the rulebook and the client's filing history — it caught errors humans had been making for years. 555 tests, penny-accurate regression, ~$0.10/filing. Live, filing every quarter for a paying carrier. |
| verdict | A skeptical AI QA agent with memory: baseline → delta runs (NEW/REGRESSED), a fact harness so every count is measured rather than composed, a signed run history its own audit made imitation-proof, flaky quarantine with expiry, and an exit-code release gate for CI. Ships verdict-qa-mcp, a read-only MCP server listed in the official MCP Server Registry. Six scored eval fixtures, misses published — including the trap it catches 1 run in 3, measured at n=3. |
| sales-agent-showcase | Clean-room showcase of a multi-marketplace commerce agent running a live resale business. Irreversible actions are human-gated structurally — the reprice call asserts its own execution grant, so there is no code path to the write without an approved approval. 30 tests, CI, runnable on synthetic data. |
| second-brain | Local-first RAG assistant: answers only from your own notes, docs and code, every claim cited to source, graded by an eval harness with citation grading. Exposed over MCP. |
| lab-control-mcp | MCP server operating a self-hosted AI lab, with a safety-gated remote shell — allowlist, no shell metacharacters, hard timeouts. |
| email-agent | Self-hosted inbox triage on local Ollama — $0 LLM cost. IMAP → classify → daily Telegram digest, scheduled via launchd, SQLite dedupe so reruns never double-notify. |
Also here: liora-studio (full-stack e-commerce + admin CMS with TOTP 2FA), greek-scythian-society-website (client site — Docker/Caddy WordPress, one idempotent script from bare VM to live), and bol-extractor (Bill of Lading PDF → structured data).
AI — agents · RAG · MCP servers · evals · agent guardrails · human-in-the-loop design · prompt engineering · Anthropic SDK · LiteLLM · Ollama · Qdrant · LangGraph Build — Python · TypeScript · FastAPI · Next.js · React · Postgres / SQLite · Docker · Playwright Quality — spec-driven development · formal test design (equivalence partitioning · boundary value analysis · decision tables) · testing non-deterministic systems · seeded-defect evals
A boundary-value pass on a client web app once caught a real off-by-one exactly at the 1 MB rollover
("1024 KB" instead of "1.0 MB") — the kind of bug nobody files a ticket for and every user notices.
An always-on lab (Mac mini + Alienware over Tailscale) running free local models behind one gateway, with MCP servers I built and a vector store for retrieval. Private by default — client data stays inside the building; I reach for a frontier model only when the task earns it.
Years of building and supporting compliance and back-office systems for small logistics businesses — IFTA filing, Bill of Lading parsing, fuel-tax reconciliation, DOT/ELD data. Most engineers never see this domain; most people in it can't build software for it. That overlap is where my best work comes from.
I'm open to AI engineering, AI automation, and QA roles — remote (US), Sacramento, or SF Bay Area. Need something built? ArtJack Technology ships websites, e-commerce, AI automation, and compliance pipelines for small businesses — fixed scope, tested, documented.
✉️ hello@artjeck.com · 🌐 artjeck.com · 💼 linkedin.com/in/evgenii-menshikov