Mateusz OzimkiewiczMateuszAI Automation · Full-Stack

I BUILD AI AUTOMATIONS AND FULL-STACK PRODUCTS

AI Automation & Full-Stack Developer · AI Product Engineer

For real business workflows — from AI agents and RAG systems to internal tools, APIs, integrations and dashboards, I design, build and validate the full product stack.

Discuss your workflowView selected work

WHAT YOU CAN COMMISSION

  • AI agents & automations
  • Internal tools & dashboards
  • Document intelligence / RAG

Selected systems

Each project proves a different kind of work you can commission — projects that are concepts or previews are labelled as such.

Flagship

Yuzuki Desktop

Local-first AI desktop with fail-closed private mode

Teams want AI inside their daily work but cannot send client files, contracts or internal notes to somebody else's API.

In private mode the conversation never leaves the machine — and if the local runtime cannot serve a turn, Yuzuki stops instead of quietly falling back to a cloud model.

Tauri · React · TypeScript · Python · Hermes Agent · Ollama

PRIVATE · LIVE-VERIFIED6 MODES · IMPLEMENTED · NOT LIVE-VERIFIEDSOURCE ONLY

View case study : Yuzuki Desktop

YuzukiSEO automation reviewPrivate · local

New conversation

  • Chats
  • Search
  • Review

You

Review this SEO automation architecture and tell me what should be automated first.

Yuzuki

Start with the repetitive handoff between lead discovery, qualification and reporting.

Recommended first automation

  1. Collect qualified company signals
  2. Enrich the account
  3. Score against the ICP
  4. Generate a review-ready brief
  5. Require approval before outreach

One ingestion job, one scoring rule set, one brief template and a review queue. Steps 01–04 can run unattended; 05 stays manual until the scoring is trusted.

Starting points

  • Explain this architecture
  • Compare two implementation approaches
  • Plan my next development task
  • Help me debug this error
  • Summarize these notes
  • Turn this idea into a step-by-step plan

Ask a follow-up question…

ContextFiles

SEO manager

  • seo-manager-brief.md
  • keyword-opportunities.csv
  • content-plan-q4.md
  • technical-seo-review.md

Automation

  • workflow-map.md
  • lead-qualification-rules.json
  • approval-policy.md

Project

  • client-requirements.md
  • implementation-plan.md
  • api-integration-checklist.md
SYNTHETIC PUBLIC DEMONo real session data
Interface demo rebuilt in HTML for this page: the application's real layout with example content. Not a screen capture, and not a captured session.

Agent

Klaus

AI orchestration and coding agent

Hand one model a whole feature and it loses the thread: context runs out, nothing reviews the result, and nobody can say afterwards what was actually checked.

Work is split by role and gated, so a step that cannot be verified is reported as unverified instead of assumed to pass.

Orchestration · Coding agent · Project context · Review gates

CONCEPT · OWNER-DEFINED ARCHITECTURE

View case study : Klaus

KlausAI orchestration & coding agent

Example task

Build an internal client dashboard.

  1. Understand

  2. Research

  3. Route

    Input
    Scoped task and context packet.
    Action
    Split the work by role and assign each role to its configured agent or model.
    Output
    Architecture, implementation and review lanes.

  4. Implement

  5. Review

  6. Validate

  7. Handoff

CONCEPTUALOwner-defined architecture · not a live agent run
A drawing of a designed workflow. No provider, model or integration is named or implied, and no stage shown has been executed.

Platform

MellyCore AIOS

Command-center architecture for shared context and agent coordination

Once several agents share one codebase, nobody can see what context they hold or what they are about to do next.

Every step is gated before anything can run, and execution stays locked in this build — there are no live providers behind it.

Context Graph · Agent Handoffs · Safety Contracts · Static preview

PARTIAL · DOCS + STATIC SLICE

View case study : MellyCore AIOS

MellyCore AIOS command center: a shared-context panel, a repository-derived topology graph, and the guarded operations workflow with execution locked.
Static preview snapshot. Execution locked; no live providers.

RAG system

RAG Document Assistant

RAG-powered document search with source-grounded answers

Support and operations staff answer the same questions over and over by hunting through procedures, contracts and handbooks.

Each answer arrives with the document and page it came from, so a reader can check the claim instead of trusting it.

Retrieval · Embeddings · Indexing · Citations

CONCEPT INTERFACE

View case study : RAG Document Assistant

RAG Document Assistant concept interface: a question, an answer written from the documents with numbered citations, and the retrieved source passages beside it.
Concept interface. Figures shown are placeholder values; no live data source is connected.

Analytics

MellyTrade

Data-dense analytics and monitoring dashboard for signals, risk and system health

Dense operational data defeats most dashboards: signals, risk, alerts and pipeline health end up on five screens nobody reads together.

One view a person can actually scan, and read-only by design — there is no broker execution behind it.

React · TypeScript · FastAPI · Analytics UI

CONCEPT INTERFACE · SAMPLE VALUES

View case study : MellyTrade

MellyTrade concept dashboard: a market overview chart, a signals list with direction, strength and confidence, and summary cards for volume, profit and loss, and risk exposure.
Concept interface, read-only. Figures shown are placeholder values; no market connection.

Other work

AI Agent Workspace

Documented multi-agent development workflow

A documented workflow connecting ChatGPT, Claude Code, Codex, Obsidian and GitHub through reusable context, task contracts and review gates.

Claude Code · Codex · Obsidian · GitHub · Documentation

AI Automation / OpsPilot

Operational signals to a human decision

A concept for a workflow I can build — not a shipped project.

Signal · Analysis · Priority · Suggested action · Human review

CONCEPT · NOT YET BUILT

What I build

The categories of work I take on, each tied to evidence on this site.

AI automation & agents
Agent workflows · Task routing · Human review · Context handoff
Evidence: Yuzuki's routing and fail-closed provider check
Document AI / RAG
Ingestion · Chunking · Embeddings · Retrieval · Source-cited answers
Evidence: RAG Document Assistant (concept interface)
Full-stack products & APIs
React · TypeScript · Python · FastAPI · Tauri
Evidence: The Yuzuki desktop product
Dashboards & internal tools
Data-dense UI · Monitoring · Risk and status views
Evidence: MellyTrade (concept interface, read-only)
Testing & validation
pytest · Playwright · GitHub Actions · Documentation
Evidence: Yuzuki's public validation suites (raw counts)

Watch a request meet the router.

In private mode it hits a hard stop instead of a cloud fallback.

PRIVATE · LIVE-VERIFIED6 MODES · IMPLEMENTED · NOT LIVE-VERIFIED

Enter the Yuzuki case studyView source (opens in a new tab)

PARTIAL · DOCS + STATIC SLICE

MellyCore

A systems architecture for agent orchestration: an operating loop where a human approves before anything is implemented. It is not a shipped public product.

Read the MellyCore case study →

MellyCore operating loopA closed loop of seven steps: observe, analyse, recommend, approve, implement, validate, record, then back to observe. Approve is the one emphasised step, labelled human in the loop.observeanalyserecommendapproveHUMAN-IN-THE-LOOPimplementvalidaterecord
  1. observe
  2. analyse
  3. recommend
  4. approveHUMAN-IN-THE-LOOP
  5. implement
  6. validate
  7. record
  8. ↺ back to observe

Implemented

  • Repository documentation and static homepage
  • Loop Operations Foundation (9 registered loops, report-only)
  • Context Gate through iteration 4
  • The static Holographic Source Arena CSS/DOM slice

Prototype

  • The legacy Live Cockpit V2

Planned

  • The full Observatory
  • Real provider adapters
  • The 3D/WebGL renderer

Simulated: model-comparison copy is deterministic local text, not live model responses.

CONCEPT · NOT YET BUILT

AI Automation

A concept for a workflow I can build — not a shipped project.

  1. Signaloperational
  2. AnalysisAI
  3. PriorityLow / Medium / High
  4. Suggested action
  5. Human review — key stagethe step that decides

How I build

  1. Discover
  2. Architecture
  3. Safety contract — key stage
  4. Build
  5. Validate — key stage
  6. Real-device acceptance
  7. Ship

Yuzuki's fail-closed provider check is the safety contract; its 23/210/63/7 suites are the validation.

Systems index

A directory view of the same work — not the navigation, just the map.

~/melly

About

AI Automation & Full-Stack Developer working across React, TypeScript and FastAPI, building AI automations and full-stack products for real business workflows — with an AI-assisted workflow built around engineering discipline: dry-run defaults, read-only modes, and honest limits.

More about me →

Have a workflow that still depends on repetitive manual work?

I build AI automation, internal tools and full-stack AI applications for businesses. Send a short description of the process — what happens today and what should happen instead — and I will reply by email.

Discuss your workflowOther ways to reach me

Colophon

Mateusz Ozimkiewicz · Available now · Poland / Remote · B2B · Freelance · Contract ·

Every claim this site makes, and how well each is proven:

  1. Local chat path, end to endLIVE-VERIFIED
  2. PRIVATE fail-closed provider checkLIVE-VERIFIED
  3. Six cloud routing modesIMPL · NOT LIVE-VERIFIED
  4. ADVANCED read-only projectionIMPL · NOT LIVE-VERIFIED
  5. Knowledge ConstellationPROTOTYPE / FIXTURE DATA
  6. Lifecycle / Job Object process treeSHIPPED
  7. Validation suitesVALIDATED
  8. Public distributionSOURCE ONLY
  9. MellyCore implemented layersPARTIAL
  10. MellyCore ObservatoryPLANNED
  11. Klaus orchestration + coding agentCONCEPT · NOT YET BUILT
  12. RAG Document Assistant interfaceCONCEPT · NOT YET BUILT
  13. MellyTrade analytics interfaceCONCEPT · NOT YET BUILT
  14. OpsPilotCONCEPT · NOT YET BUILT