Solo-built software platform running a car servicing business's daily operations — replaced paper notebooks and Excel entirely. Mechanics record jobs, look up vehicles via the DVLA API, and track customers through a Telegram bot; completed jobs auto-generate and send invoices through FreeAgent. An AI extraction pipeline processes 500+ supplier invoices per year — PDFs arrive via a Gmail Apps Script or direct Telegram upload, get deduplicated, and pass through a multi-pass LLM pipeline that extracts accounting fields into PostgreSQL. The public-facing site is trilingual (English/Ukrainian/Russian), built with progressive enhancement so it works fully without JavaScript. Cut monthly bookkeeping from 10 hours to under 30 minutes.

Freelance contract platform where AI agents autonomously verify whether GitHub milestones have been completed before releasing payment. Per-issue verification agents run in parallel, each with its own browser session — they inspect repository state via the GitHub API and navigate live deployments through Playwright to gather evidence. A Claude judge reviews the full evidence trace and can override any verdict it finds insufficiently supported. Verified milestones trigger a real Solana escrow transfer with an on-chain Memo receipt linking payment to the specific milestone and report.

AI grading platform where teachers upload a mark scheme PDF and the system extracts structured rubrics — question types, mark allocations, and criteria — then grades student submissions against them. Each graded submission returns per-criterion feedback with evidence quotes pulled from the student's work. Supports batch grading with up to 5 concurrent AI calls across multiple providers (Claude, Gemini), with the frontend polling for completion in real time. Includes an anti-prompt-injection rule to prevent students from gaming the grader through their submissions.

Natural language CLI assistant for Linux that translates plain-English requests into bash commands and executes them. The LLM responds only in structured JSON with one of three action types — execute (run commands), analyze (gather information and re-plan), or tool call (invoke from a registry of 7 specialized tools including system diagnostics, web research, and image fetching). Destructive commands are pattern-matched and gated on user confirmation. A self-correction loop feeds errors back to the LLM and resets gracefully after 4 consecutive failures rather than retrying indefinitely.

Unlimited multilingual Wordle clone supporting 20 languages including 9 non-Latin scripts (Arabic, Armenian, Bengali, Georgian, Greek, Hebrew, Korean, Telugu, Urdu). Standard JavaScript string methods break on multi-codepoint characters, so all character counting and validation uses grapheme-aware splitting throughout — from word filtering on the server to guess validation and keyboard input on the client. Each language gets a filtered on-screen keyboard derived from its locale layout. Game configuration (language, word length 2–8, attempts 4–8) is serialized into the URL for sharing.

Data pipeline that joins word frequency data (OpenSubtitles, ~70 languages) with UniMorph morphological tables to produce POS-tagged, frequency-ranked word lists across 40+ languages. The two upstream sources use different ISO standards (639-1 vs 639-3), so the pipeline bridges them with a reference table and manual overrides. An inflection-backtracking mode maps surface forms back to canonical lemmas so frequency signal pools onto dictionary entries rather than conjugations. Powers the word banks for Wordle International.