Move Make scenarios and Zapier zaps to n8n: an importable workflow plus a report on every step.

Stack: TypeScript · Node.js · MCP · n8n

• every step ends up converted, to review (one named thing to check) or manual (left in place, disabled, with a note on how to finish it)
• routers, filters, iterators, formulas and error handlers carry over; nothing is dropped
• works from Claude, Cursor, VS Code or the terminal; scenarios never leave your computer

Result: 1,432 public Make scenarios and 148 zaps converted with 0 broken outputs; 83% of Make steps convert automatically or with one named check.

Try it: npx migrate-to-n8n convert my-scenario.blueprint.json

_{329 tests · on npm and in the official MCP Registry · imports and live webhook runs checked on real n8n 2.x and 1.x · my own open-source tool}

Ukraine's public procurement (Prozorro) for an AI assistant or the terminal: open tenders, upcoming plans, market, competitors, red flags.

Stack: Python · MCP · SQLite · Prozorro API

• syncs only your oblasts from the public API, newest tenders first, resumes after a break, no keys
• unit prices, win rates, single-bidder share: computed by code, never by the model
• red flags with evidence; evals: precision and recall 1.0 on 35 cases, totals reconciled with raw data

Saves: a bidder gets new tenders and next month's planned purchases every morning instead of browsing prozorro.gov.ua by hand.

_{228 tests · on PyPI and in the official MCP Registry · one-click desktop extension · CSV for Excel without AI · my own open-source tool}

Topic in, finished documentary out: script with sources, voice-over, shots, mastered mp4 with subtitles and a thumbnail.

Stack: n8n · TypeScript · Python · FastAPI · FFmpeg · OpenCV · Postgres · Docker

• every shot is checked (black, frozen, wrong length, face drift); 3 tries, then a safe fallback
• picture and narration stay in sync; −14 LUFS like broadcast
• runs end to end without API keys in demo mode

Saves: a documentary like this takes 15–20 h by hand (est.: script 4 h, voice 1 h, footage 4–6 h, editing 6–8 h); here it is ~1 h of review.

_{75 + 92 tests · e2e on a real n8n: topic → video in ~3 min}

Ask a data warehouse in plain English; get a table, a chart and the trace of how the answer was found.

Stack: Go · Fiber · React · TypeScript · Postgres · BigQuery · OpenRouter

• the model acts only through tools the user's role allows; every call is audited
• every SQL query goes through a validator; expensive ones need a click to approve, with the price
• answers with numbers no tool returned are rejected

Saves: a manager gets an answer in seconds instead of waiting for an analyst.

_{97 Go + 12 web tests · 10/10 e2e · reference version of the AI analyst I built on the client's BigQuery platform}

A support agent that answers only from its knowledge base and hands everything else to a person.

Stack: Python · RAG · BM25 / Qdrant · Chatwoot

• personal data is caught before any model sees the ticket
• billing, VIP and abuse go straight to a person; unsure → hand-off with a ready note
• measured on a golden set and an adversarial set, in CI, without keys

Result: 0 wrong answers on the golden set, 62/62 adversarial tickets handled.

_{130 tests · client system: ~30% → ~70% of tickets}

Jira sprint metrics computed by code and written up by an LLM, as a CLI and an MCP server.

Stack: Python · MCP · Jira API · evals · Langfuse (optional)

• story points, cycle time, time in status: computed, never guessed by the model
• a guard rejects any number the model did not get from the code
• evals measure the guard: precision 1.0, recall 0.79 on 22 cases

Saves: team leads get the sprint report without collecting numbers by hand.

_{90 tests · client system: 8 team leads, 9 departments}

n8n workflows as TypeScript: tested Code-node logic, deterministic builds, a deploy that cannot break production.

Stack: TypeScript · n8n Workflow SDK · esbuild · n8n API

• deploy saves a draft, then publishes; a UI edit stops the deploy instead of being overwritten
• rollback with a backup; dependency-ordered deploys
• verified end to end on a real n8n 2.40

Saves: no more hand-edited production workflows; one command deploys and checks all of them.

_{70 unit tests · 8/8 e2e checks}

Several role-based agents, each on its own model, argue in rounds until they all accept one answer.

Stack: Python · OpenRouter · aiogram (Telegram)

• architect, critic and pragmatist answer on their own, then review one draft in rounds
• the moderator accepts or rejects every objection; open ones are reported as dissent
• one decision with reasons in Telegram; the full debate as a file

Saves: a second and third opinion on design decisions in minutes.

_{77 tests · my own tool; the original code was lost, rebuilt for publication}

My own desktop messenger: end-to-end encrypted chats, voice and video calls, screen sharing, built-in mail.

Stack: TypeScript · Electron · React · Fastify · PostgreSQL · libsignal · LiveKit

• E2E encryption on the Signal protocol
• its own mail service; installers built in CI with auto-update
• beta with a small group of users

_{~28k lines · 363 tests · built solo}

AI: LLM agents and tool use · RAG · MCP · evals and guardrails · OpenRouter, Claude and OpenAI APIs

Automation: n8n (workflows as code) · Make · Zapier · migrations to n8n · Slack and Jira apps · Telegram bots

Engineering: Python · TypeScript · Go · FastAPI · Fiber · React · PostgreSQL · SQLite · BigQuery · Redis

Delivery: Docker Compose · Linux VPS · Nginx / Caddy · GitHub Actions

✉️ kotiniiqq@gmail.com · 💬 Telegram @kotiniiqq
📍 Antalya, Turkey · remote · English B1